Solar Grid Optimizer

A physics-informed neural network model that can predict solar irradiance, and calculate the most optimal replacement position for solar panel anywhere in the world from outputs.

Comment

Inspiration

Even though there are numerous advantages of solar energy, it’s also important to optimize the efficiency of solar panels when it comes to utilizing solar energy, and sometimes the optimization process can be difficult. The reason for this difficulty is the multi various factors that can affect the most optimal efficiency taken from the solar panel. Alternatively, using conventional computation techniques can be beneficial for simple optimization problems, but in real world application, the results could be deviated because of their strictly adherence to constraints. Therefore, we put forward an innovative approach to these problems by including physics-informed neural networks (PINN).

What it does

Solar Grid Optimizer is a PINN model that predicts irradiance and obtains the efficiency values of solar panel replacement from them. This model makes complex computations of the differential equations effectively by training a network on the solar irradiance. With this project, we observe how innovative the integration of neural networks can befor optimization problems, including energy optimization. In order to improve renewable energy usage, we provide users with the most optimal efficiency values based on their location, their timeline and specific features for solar panel replacements; in this way, users won’t need to waste time to find out the best location, and earn more money with specific recommendations from the model about replacement position of the panel.

How we built it

Our model encompasses general neural network architecture, and it is built with one input layer with 6 neurons; 3 hidden layers, two of them 128 neurons, the other 256 neurons; one output layer with 1 neuron. However, Unlike traditional artificial neural networks, these neural networks are trained with governing equations which enable the model to learn complex physical computations accordingly. Thanks to this specific feature, they can be adaptable to any field in physics. In our case, the PINN model is trained to predict solar irradiance under considerations of world conditions, including several factors that affect the calculation of irradiance.

Challenges we ran into

During the built of this model, we primarly had a hard time finding out a way how to approach it. This difficulty was because of complexity of the model. We delved into the deep analysis for solar irradiance computation, and engaged with sophisticated processes. Especially, determining the related factors that significantly effect the solar panel efficiany was way hard. Some of the factors were negligible; some of the factors were substantial, so we had to process this period rigirously.

Accomplishments that we're proud of

The biggest accomplishment is actually to build the main strategy for overall structure because the PINN integration plan for renewable energies, spesifically solar energy, was way harder than we thought. Another important accomplishment is that we were able to train our model over 96% accuracy which shows how accurate its predictions are compared to the physics based results.

What we learned

We learned how engaging and innovative physics-informed neural networks stand out as solutions for sustainable development goals, particularly renewable energy. Thanks to this model, we're sure that such processes can be made efficiently since its flexibility is really remarkable. Additionally, we also learned that the optimization problems are one of the most fundamental parts in these real world problems, and innovative solutions are necessary for global development.

What's next for Solar Panel Optimizer

Our main strategy with Solar Panel Optimizer is to expand its coverage in solar physics. We aim to include other factors that make the results much closer to the real world results. Additionally, on the side of solar panel replacement, we intend to broad the related subjects; thus, the model will not only be responsible for solar-based computation, but it is also trained to predict optimization of wind turbines.

Built With

Share this project:

Updates