Challenge 03 — Climate Optimizer

Problem Statement

We know roughly what needs to happen to hit net zero. The hard part is working out exactly how — and when — to make it happen.

Energy grids must balance supply and demand in real time, but renewable output is intermittent and hard to predict. Buildings account for roughly 40% of global energy consumption, yet most operate on fixed schedules with no awareness of what the grid is doing. EV charging is growing exponentially, but uncoordinated charging creates demand spikes that push grids back to fossil fuels at precisely the wrong moment. In each case, the data to make better decisions already exists — the missing piece is a system intelligent enough to use it.

The challenge is to use Predictive AI to find the optimum. Take messy, real-world data about energy systems, carbon output, and demand patterns and produce a recommendation that is faster, broader, and more accurate than any human planner working alone.

Data Available

The following datasets are suggested as starting points. You are not limited to these — use whatever sources best support your solution.

• Hugging Face — Energy & Forecasting Datasets — Energy consumption, demand forecasting, and grid datasets ready for use with ML models.
• ENTSO-E Transparency Platform — European electricity grid data: generation by fuel type, load forecasts, and actual demand — updated in near real time.
• Electricity Maps — Real-time and historical carbon intensity of electricity grids across 50+ countries.
• Open Power System Data — Harmonised European electricity system data including renewables generation and national load profiles.
• NREL — National Renewable Energy Laboratory — US solar and wind resource data, energy system modelling tools, and EV charging behaviour datasets.
• Copernicus Climate Change Service — European climate reanalysis and seasonal forecasting data for renewable energy yield prediction.
• EU Building Stock Observatory — European building energy performance data including consumption benchmarks and renovation rates.
• Swiss Federal Office of Energy (SFOE) — Swiss energy statistics covering electricity generation, consumption by sector, and renewable capacity.
• opendata.swiss — Energy — Swiss cantonal and municipal energy data including smart meter consumption profiles.

Data Integration

Your solution must demonstrate that predictions or recommendations are grounded in real data — not generated from model assumptions alone. Data integration is intentionally broadly defined and includes any of the following:

• Pulling live or historical data from an API and feeding it into a model
• Training or fine-tuning a forecasting model on a downloaded dataset
• Using structured data to constrain or validate model outputs
• Combining multiple datasets to build a richer feature set

A team that downloads a grid carbon intensity CSV, builds a simple forecasting model, and uses it to recommend an optimal EV charging window is fully meeting this requirement. What matters is that the output is traceable to real data.

Expected Output

• Predictive Model or Optimisation Engine — A working implementation (e.g. a time-series forecasting model, a scheduling algorithm, or a reinforcement learning agent) that produces a concrete recommendation from real data.
• Insight Dashboard — A visual output showing the model's predictions and the impact of its recommendations (e.g. "Shifting EV charging to 02:00–05:00 reduces carbon emissions by an estimated X kg").
• Approach Summary — A brief explanation of your method: what you are predicting or optimising, what data you used, and how you validated it. No specific algorithm required.
• User Scenario — A clear demonstration of who would use this tool and what decision they could make with the output (e.g. a building facilities manager, a grid operator, a fleet manager).

Demonstrating Reliability

Predictive models fail when conditions fall outside their training data — an unexpected heatwave, a sudden drop in wind generation, an anomalous demand spike. Your presentation must include a live demonstration of at least one of the following:

• The model displays a confidence interval or uncertainty range alongside its prediction, rather than presenting a single number as fact
• The model flags when input conditions are unusual or fall outside the range of its training data
• The model is shown handling an edge case and produces a sensible, conservative response rather than a confident wrong answer

This will be evaluated by judges during the 5-minute presentation.

Success Criteria

• Prediction Quality — The model produces outputs that are plausible, data-grounded, and accompanied by an honest assessment of their reliability.
• Actionability — The recommendation is specific enough for a real decision-maker to act on — not just a trend visualisation.
• Data Efficiency — Teams are judged on how effectively they used available data, not model complexity. A well-applied pre-trained forecasting model is valued equally to a custom architecture.
• Impact Potential — The degree to which the tool could realistically reduce carbon emissions or improve energy efficiency if deployed at scale.