The BTP tool draws inspiration from the Cal ISO Day-Ahead Outlook to create a site level and portfolio level forecast of expected demand and expected on-site generation.
What it does
The BTP tool is a predictive image of a PV+Storage system. The tool accesses, home power usage, battery state of charge and PV array production data against forecast weather conditions at the site to generate a forward looking asset performance profile. This profile can be used by system owners (like sunrun) to map battery dispatch schedules for grid services while assuring individual customers always have backup availability in case of outages.
How we built it
Our team has built an API to consume Dark Sky's weather forecast database. A key function converts weather data into power production, and will act as a the predictive engine for this system. We have built a code that reports PV performance range against average daily cloud cover, and will use this data to create forward looking Available Power figures for energy storage assets.
Challenges we ran into
The first challenge is the cleanliness of the dataset. The 19 million data-points was raw and not well arranged. Also, there were no readme files provided to describe the units and the meanings of data attributes. How to generate forward looking PV performance data against weather. We are concerned that our margin of error increases with increasing cloud cover(trend spotted as per out model). We built our model based on a single house, more time would have encouraged us to try out different scenarios. We did not have time to fully develop the predictive, machine learning component of this tool, but we have been able to develop the thought models that will influence the final iteration of this product.
Accomplishments that we're proud of
We have created a good preliminary backend scheme for this tool, drawing from several databases based on site data input by the user. We have managed to even find a co-relation between the Cloud Cover percentage and the mean error range out of the KNN model.
What we learned
It is important to bound this tool in terms of information that is valuable to the asset owner, grid operator, and customer.
What's next for Blue Team Party!
Honing the system to provide Asset Owner level and Customer level views.
Adding control functionality so that monitored assets can be throttled remotely. Adding more robust predictive models using ML. (Like Artificial Neural Networks, Decision Tree Methods)