Solar Mirror
Building Spatial Augmented Reality for Explainable Robot Vision
2023
For: Harvard GSD SCI 6484: Data Science for Environmentally Responsive Buildings
Role: Grasshopper setup, data collection, classifier training
Partner: Ya Qin (Harvard)
Overview
Natural lighting effectively reduces building energy consumption and enhances occupant well-being. However, its potential in buildings is often underutilized, particularly during the long winter months with limited daylight and the summer months at high sun altitudes, presenting significant opportunities for daylight redirection (DR) technologies and adaptive architectural designs. This research proposes a DR strategy using an automated multi-angle solar mirror, dynamically controlled by a machine learning model, to reflect desired daylight indoors. By controlling the mirror's pitch and yaw according to the sun's altitude and azimuth, the system can extend natural lighting hours, reducing reliance on artificial lighting and optimizing indoor lighting quality and comfort.
Problem
In locations with relatively higher latitudes, the daylight hours and solar zenith angles vary significantly between seasons (González, Da Costa, Tam, & Haddad, 2024). In the context of the built environment, composed of the adjoining and stacking of basic architectural forms of rectangular boxes, the seasonal daylight variations directly affect indoor energy costs, environmental quality, occupant comfort, and the human circadian system (Roenneberg et al., 2007). The potential of natural lighting is often underutilized, especially during the long winter months with limited daylight hours, and summer months when the sun is high in the sky , which makes sunlight harder to reach indoors due to the steep angles or causes glare due to increased intensity.
In order to systematically explore the effectiveness of the proposed DR strategies and to achieve a thorough understanding of the relationship between indoor lux and the sun’s position throughout the year, this research adopted a combination of methodologies to discover a more data-driven approach to managing indoor lux level. Inspired by data-driven approaches to optimizing indoor lighting, such as the work of Doe & Smith (2016) on automated blinds, this study applies a similar methodology to managing indoor lux levels. We began by assessing occupancy behavior and user preferences for indoor lighting, followed by geometric modeling in Rhino using a generalized shoebox model to evaluate daylight utilization. Since the study aims at precise spot lux improvement, sensors for lux detection and analysis were evenly distributed on top of a round table with a diameter of 1.5 meters. Parametric design techniques were applied to optimize mirror designs for light redirection by adjusting variables such as time of day, mirror tilt, and rotation. Simulations were conducted using ClimateStudio and Rhino Grasshopper to generate illuminance data, with lux sensors distributed over a controlled space. To support the analysis of our generated data, machine learning models—including Linear Regression, Random Forest, and Multilayer Perceptron—were trained and validated using normalized data. The models' performance was compared using Mean Squared Error (MSE) and prediction graphs. The next step of this research will involve utilizing transfer learning to extend the study to a wider range of geographic locations and diverse interior configurations, including variations in window shapes and positions. This approach will allow the model to adapt its understanding from previous environments to new, yet related, scenarios.
Shoebox Model Approach
Shoebox models are simplified building energy simulations commonly used in architectural energy modeling, particularly during the early design stages. This approach offers several advantages:- Simplicity and Speed: Given the diverse parameters of different buildings, shoebox models are the simplest and most efficient way to represent building performance. By reducing complex geometries to simple box-shaped zones, they significantly shorten simulation times compared to detailed whole-building models. (Battini, Pernigotto, & Gasparella, 2021) This allows for rapid iterations and the comparison of design alternatives early in the process.
- Focus on Key Parameters: Shoebox models enable us to concentrate on essential aspects of the design. With their simplicity, we can quickly run multiple simulations in a parametric manner to identify the most effective design, maximizing our ability to explore various options efficiently.
- Retaining Important Parameters: Despite their simplicity, shoebox models still maintain crucial parameters for simulating indoor illuminance. Using tools like ClimateStudio, we can simulate accurate lighting and material interactions, ensuring that key environmental factors are not compromised.
a) Shoebox model and table position. b) Top view with light simulation without mirror. c) Shoebox model with mirror.
Realworld Site: 40 Kirkland East facing window
Finding the right mirror size and position
Data Collection
The focus of this exploration is to test different mirror design variations and their effects on indoor lux. Polished silver with a reflectance rate of 95.98% was assigned to the mirror in the simulation software. For daily simulations, data were generated every fifteen minutes from 9:00 AM to 5:15 PM. Preliminary indoor illuminance variation patterns and the differences in lux optimization potential among various mirror designs and interior materials were studied and compared using all-day data from July 1st. July, in Boston—where the initial model was based—had the lowest average indoor illuminance. This preliminary sample study provided an efficient and accurate understanding of the daylight utilization potential of different designs, allowing the research to proceed to the annual data study in the next step. Simulation results for each scenario were compared against a “no mirror” baseline study.Parametric Control Examples: a) rotation 10°; tilt 10°. b) rotation 10°; tilt 20°. c) rotation 20°; tilt 20°. d) rotation 20°; tilt 40°. e) rotation 30°; tilt 30°.
Python Script automating the data collection process
Model Training
Machine learning is increasingly being utilized in smart buildings for its ability to enable buildings to respond to environmental changes and manage energy consumption and indoor comfort levels (Cheng, Zhang, Chen, & Liu, 2021). In the development phase of this study, the focus is on understanding the relationship between input and output data calibration and evaluating machine learning models to maximize and refine the performance of the DR technology. The input data, commonly referred to as “X features” or “input variables,” include the hinges, rotation angles, and tilt angles of the mirror, as well as the azimuth and altitude of the sun. The rotation and tilt angle data were generated through simulations in ClimateStudio, and the sun’s azimuth and altitude data were sourced from the data service website of the United States Naval Observatory (USNO) (U.S. Naval Observatory, n.d.). The USNO is the national research facility of the U.S. Navy responsible for astronomical data. The output data, or the values the models aim to predict—referred to as the “Y predicted output”—was the mean lux of the targeted working plane in the shoebox model.Linear Regression, Random Forest, and Multilayer Perceptron (MLP) models were tested. These are commonly used machine learning algorithms with different predictive modeling techniques for pattern discovery and categorization. Each model has its own strengths and applicable situations. Linear Regression is efficient when the input variables and output values have a linear relationship. Random Forest is an optimal ensemble model for handling complex features, aggregating results from multiple decision tree predictors (Breiman, 2001). MLP is the only neural network, or deep learning model, among the three. It consists of multiple layers of neurons, where each layer processes its input and passes it to the next layer until the final output layer makes a prediction or classification. It is particularly effective for handling high-dimensional data (Sarker, 2021).
The purpose of testing these models was to understand the underlying relationships among the values based on the quantitative performance feedback provided by each model. The ultimate goal of this phase is to leverage machine learning models to establish the relationship between input data and controllable daylight output, enabling occupants to optimize lighting conditions and enhance energy efficiency through the use of reflective mirrors.
The purpose of testing these models was to understand the underlying relationships among the values based on the quantitative performance feedback provided by each model. The ultimate goal of this phase is to leverage machine learning models to establish the relationship between input data and controllable daylight output, enabling occupants to optimize lighting conditions and enhance energy efficiency through the use of reflective mirrors.
Normalized Model Performance Assessment
The MLP was initialized with four hidden layers containing 256, 128, 64, and 32 neurons, respectively, and was trained over 200 epochs. Based on the MSE test values, the MLP proved to be the best-performing model for this research.