Power Laws: Cold Start Energy Forecasting

Usually with buildings, bigger historic datasets yield more accurate consumption forecasts. The goal of this challenge is to provide an accurate forecast from the very beginning of the building instrumentation life, without much consumption history. #energy

€23,000 in prizes
Completed oct 2018
1,289 joined

alt-text

Each challenge in the Power Laws series explored a different aspect of energy efficiency and management. The winning algorithms were released under an open source license to spread understanding about how energy modeling works and what approaches are most effective.

Why

Increasing the efficiency of energy consumption has benefits for consumers, providers, and the environment. Good forecasts of building energy use play a critical role in planning efficient policies, optimizing storage for renewable sources, and detecting wasteful anomalies.

But what if a building is just becoming operational, and we don't have much past data for making forecasts?

The Solution

This “cold start problem” was the focus of a data science challenge that drew on data from Schneider Electric. More than 1,200 data scientists competed to build the most reliable predictions given only a few days of historical data for each building. What’s more, competitors had to work with different time windows to provide accurate hourly, daily, and weekly forecasts.

The Results

The top algorithm brought down the median absolute percent error across daily and weekly forecasts to 7% - less than half that of the LSTM benchmark.

Overall the winning approaches carefully considered how to combine the limited historical information provided with useful meta-data (like holidays), while tailoring different models for each time window (hourly, daily, weekly) in order to build the most reliable forecasts.


RESULTS ANNOUNCEMENT + MEET THE WINNERS

WINNING MODELS ON GITHUB

DATASET ON SCHNEIDER ELECTRIC EXCHANGE