SolarSight: Deep Learning-based Solar Energy Prediction Tool

Inspiration

Renewable energy is critical for reducing global carbon footprint, with 70% of all energy coming from non-renewable sources. Climate change is inevitably catching up and renewable energy must become the energy of the future. However, balancing solar energy with other energy sources remains difficult, due to fluctuations in production levels. With this in mind, I developed SolarSight, a Deep Learning-based tool to predict solar energy output ahead of time.

What it does

Using Deep Learning neural networks, SolarSight can take in sensor data (radiation, wind speed, humidity, etc.) and predict the energy output of a solar energy grid for the next hour. With just 9 months of training data, SolarSight had a mean absolute error of less than 1.5%.

How I built it

SolarSight uses Neural Basis Expansion Analysis for Interpretable Time Series Forecasting (N-BEATS) architecture with 4 hidden layers, with 256 neurons per layer. SolarSight's model takes input features including wind speed, air pressure, sunshine, radiation, and relative humidity, and estimates the output power from the solar grid in megawatts. The chosen coding language was Python with Tensorflow, and the IDE used was Google Colaboratory.

Challenges I ran into

The biggest challenges included debugging the code to ensure SolarSight's model was working correctly, including fine-tuning model parameters and hyperparameters. In addition, data wrangling was a difficult process to understand how to adjust the features for the model to train effectively.

Accomplishments that I'm proud of

Discovering firsthand the versatility and power of machine learning in a truly applicable circumstance was a very new experience for me. SolarSight proved to be much more accurate than expected, with an error margin of less than 1.5% given just 9 months of training data. Working on SolarSight provided me with valuable experience in how to develop and train effective models to be used in the real world.

What I learned

Learning TensorFlow and using the N-BEATS architecture were both new experiences for me, which involved countless hours of trial and error work on code. Beyond learning how to code using TensorFlow, and developing neural network models, I also learned how to visualize data to gain intuition on the information I am dealing with.

What's next for SolarSight: Deep Learning-based Solar Energy Prediction Tool

I aim to look into other model architectures to develop an even more accurate model so SolarSight becomes a standard tool in the energy industry. I also aim to develop a more generalized version of SolarSight's model, by training on different types of solar grids and their respective outputs so any solar grid can use the same model for accurate results. Putting SolarSight into the hands of everyday consumers is a goal I am working towards, so developing a GUI for both industrial and consumer use is a future goal.

Built With

• deep-learning
• google-notebook
• machine-learning
• neural-network
• python