Weekly Workflow: Use LunchBox ML to Predict Solar Radiation

The process begins by creating or sourcing a dataset that can be used to train a predictive algorithm. For this example, we use Ladybug to calculate Incident Radiation for a series of planes.

1. Start a session of Grasshopper.
2. Choose a geographic location and use Ladybug’s Download EPW and Import EPW components to bring weather data into Grasshopper. An EPW file, short for “EnergyPlus Weather File,” is a text-based, comma-separated file format used to store weather data for building simulations. These files contain hourly weather data, including temperature, humidity, solar radiation, wind, and precipitation, typically for a full year. EPW files are widely used by energy modelers and building consultants for simulations involving daylighting and energy performance analysis. Ladybug includes components that direct you an interactive map where you can find links to EPWs that can be used in Grasshopper.
3. Use the data from the EPW as input values for Ladybug’s Sky Matrix component, which calculates radiation for each part of the sky at the given location.
4. Create a series of planes where radiation will be calculated. This example creates a 3D array of points, and then creates a series of rotated planes around the points, so that the dataset will represent incremental 360-degree surface rotation. Then use this geometry run Ladybug’s Incident Radiation component.
5. Evaluate each plane to find their Normal vector, then calculate each vector’s Angle from the X, Y, and Z axes as separate lists.
6. Organize the Incident Radiation Results (kWh/m2), Angle from X, Angle from Y, and Angle from Z lists and use the ExcelWrite component to save this data as an Excel spreadsheet.

The next step is to use the dataset you created to train an algorithm. In this case, we will use a Machine Learning (ML.Net) Regression Trainer. This is a category of trainers that is used for predicting continuous values (like price, temperature, etc.). Regression trainers are used when the output values are real numbers, as opposed to a category.

We’ll configure the Regression Trainer as a Light GBM (Light Gradient Boosting Machine) type. This is a specific trainer in ML.NET that is based on a high-performance, tree-based algorithm that works well for Regression and Classification problems. The decision tress it creates help the algorithm to “learn” in successive steps, so that it does not need to start from scratch when an error is encountered — it can simply return to the previous branch in the tree and move forward.

To summarize this, Regression Trainers are a class of ML training that is good for predicting values along a continuous domain, and a Light GBM trainer is a type of Regression Trainer that learns from errors and improves at each step.

1. Create a new Grasshopper document
2. Use an Excel Reader component to connect to the Excel file that was created in Part 1 (above).
3. Use the “Create Data Set” component to organize the data from the spreadsheet into Labels (the column headers) and values.
4. Add LunchBox’s “Light GBM Trainer” component to the canvas.
5. Feed the data set values into LunchBox’s Regression Trainer component. The AngleX, AngleY, and AngleZ fields should be connected as “Features”. The kWhM2 field should be connected as the “Label”. Connect the Light GMB Trainer to the “Trainer Type” input.
6. Right click on the center of the Regression Trainer component and save the regression trainer as a ZIP file. See our documentation on how to save trained models here.

In this step, we’ll use the model from Part 2 to predict Incident Solar Radiation on a series of panels.

1. In a new Grasshopper file, create a panelized surface, and evaluate each of the panels to find their Normal vector. Then, use the same workflow from Part 1 to break the vector into it’s X, Y, and Z angles.
2. Load the trained model from Part 2 into the document.
3. Use the “Regression Tester” component to predict the Incident Solar Radiation values for each panel.

LunchBox ML includes grouping components, such as Gaussian Mixture, that help users to organize elements from a given list into groups their most similar elements. This can be a helpful utility to break down complex information into smaller subsets, and in the case of grouping similar geometry can be used to facilitate the design of modules for construction.

1. Define numerical panel characteristics such as Incident Solar Radiation, Area, Out of Plane Distance, and point-to-point dimensions.
2. Structure this into a dataset. For this example, we used the “Entwine” and “Flip Matrix” components.
3. Feed the values into a “Gaussian Mixture” component and define the number of groups.
4. Visualize the groups by color. In this example, a gradient is used to set each group as a different color.