Climate change is one of the defining challenges of our time. While governments, companies, and researchers are all working toward solutions, innovation is also happening among students who are beginning to explore how technology can contribute to cleaner energy systems.
In a recent conversation with students Esha Bhatnagar and Katelyn Thilak, I learned about their project exploring how machine learning can be used to predict solar panel efficiency. Their work combines data science with a physical solar experiment, demonstrating how computational models and real-world testing can complement each other when studying renewable energy systems.
From Climate Concern to a Research Project
For Esha and Katelyn, the project began with a broad concern about climate change. They described it as a global issue that affects the entire human population, motivating them to think about potential solutions.
When considering where they could contribute, they focused on energy production. Fossil fuels remain one of the largest sources of greenhouse gas emissions, so they chose to investigate solar energy as a cleaner alternative. Their goal became developing a way to predict solar panel performance using machine learning, helping improve how renewable energy systems are managed.
Building the Foundation: Data and Machine Learning
The students began their project by collecting open-source datasets related to weather conditions and solar energy production. The data included variables such as solar irradiance, temperature, cloud coverage, wind speed, AC, and DC power output.
Using Python and machine learning libraries, they loaded and preprocessed the data. This included converting timestamps, organizing relevant variables, removing missing values, and capping outliers.
They then split the dataset into 80% training data and 20% testing data before training multiple machine learning models. The models they tested included: a physics-based baseline model, linear regression, gradient boosting, random forest, and neural networks. To evaluate the models, they used metrics such as MAE (mean absolute error), RMSE (root mean square error), and R2 to measure prediction accuracy and determine which algorithm performed best.
Their initial baseline model achieved about 73.8% accuracy using R2, which served as a reference point when evaluating the more advanced algorithms.
Comparing How Different Models Handle Data
One of the most interesting parts of their project was comparing how different machine learning models interpreted the same data.
The linear regression model assumed relatively simple relationships between variables. In their implementation, they used a second-degree polynomial equation to approximate how factors such as weather conditions affected energy output.
However, the students found that real-world environmental variables often behave in complex ways. For example, wind does not always have a consistent relationship with energy production. Moderate wind can cool panels and improve efficiency, while severe wind may cause structural damage and disrupt energy output.
Because of these complexities, models such as random forest and gradient boosting were better able to capture non-linear relationships in the data. These models analyze patterns through multiple decision trees, allowing them to detect more complex interactions between variables. Ultimately, gradient boosting and neural networks produced the strongest predictive results in their experiments.
Challenges Along the Way
Like most research projects, the team encountered several obstacles.
One major issue involved data leakage between AC and DC power variables. Initially, the model began using DC power to predict AC power instead of relying on weather data. This meant the model was not actually solving the intended problem. To fix this issue, they removed DC power from the model inputs, forcing the algorithms to rely only on environmental variables.
They also faced difficulties while working on the physical experiment. Because the solar panel produced very small amounts of power, it was difficult to obtain stable measurements. Environmental factors such as room lighting and object movement could affect the readings.
To reduce interference, they often conducted their tests at night so that outside light would not influence the experiment.
Bringing the Model Into the Real World
Beyond the computational models, the students built a physical solar panel experiment to test their predictions. Their setup included a small solar panel, a desk lamp acting as a light source, and a multimeter to measure energy output. They then simulated different environmental conditions: cloud cover by placing paper over the lamp, wind using a fan, and changes in light intensity by adjusting the lamp.
By comparing the results from their physical experiment with the predictions generated by their machine learning models, they were able to visualize how well the algorithms matched real-world behavior. Although the results were not identical, the graphs showed that the physical measurements and model predictions followed similar trends.
A Surprising Finding About Temperature
One of the most interesting discoveries from their project challenged a common assumption about solar energy.
Many people assume that hotter days produce more solar energy. However, the students found that temperature actually has a negative relationship with solar panel efficiency.
Higher temperatures increase electrical resistance within the panel’s components, reducing energy output. As a result, cooler sunny days may actually produce more electricity than extremely hot days.
Potential Real-World Applications
The students believe their approach could have several real-world applications.
More accurate solar forecasting could help power grid operators better balance electricity supply and demand. If operators know how much energy solar panels are likely to produce, they can plan energy storage and distribution more effectively. Improved forecasting could also help reduce reliance on fossil fuels as backup energy sources.
Their research could also contribute to improvements in solar panel systems themselves. For example: optimizing panel placement and tilt angles, developing automatic cleaning systems to remove dust accumulation, and designing adjustable bases that maximize sunlight exposure
What Comes Next
If they had more time and resources, the students say they would expand the project in several ways.
First, they would collect more real-world data from locations beyond the United States, allowing the model to be tested under different environmental conditions. They would also experiment with additional machine learning algorithms beyond the five models they tested.
On the physical side, they would like to build a more advanced experimental system that simulates the sun’s movement across a full 24-hour cycle. This could help better replicate how solar panels behave in real-world environments.
Another idea is to build an automated solar tracking system that adjusts the panel’s angle to follow the sun throughout the day.
Advice for Students Interested in Climate Tech
When asked what advice they would give to other students interested in climate research or technology, both emphasized that solving climate change requires contributions from many people and many ideas.
Machine learning, engineering, and other emerging technologies are rapidly evolving fields. By exploring how these tools can address environmental challenges, students can help develop new solutions.
As they put it, no single idea can solve the problem alone, but collective innovation and experimentation can move us closer to meaningful change.