nathanielsimard/solar-irradiance-prediction

This project predicts how much solar energy will reach specific ground stations by analyzing satellite imagery over time. Instead of just looking at a single weather snapshot, it uses sequences of satellite images taken 30 minutes apart to forecast solar irradiance, a key metric for solar power generation and grid management. The core idea is straightforward: satellite images contain clues about cloud cover, atmospheric conditions, and weather patterns that affect how much sunlight reaches the ground. The project trains machine learning models (called Conv2D and Conv3D) that learn to recognize patterns in these image sequences and predict solar energy levels at seven different weather stations across a region. The images come in multiple channels, different types of satellite data, giving the model a richer picture than color photos alone would provide. Each image is 650 by 1500 pixels, and the model looks at 10 consecutive images to make its prediction. The practical use case is straightforward: solar farms and energy companies need to know how much power they'll generate minutes to hours ahead. A utility operator could run this model to forecast whether a solar installation will produce peak power or drop due to cloud cover, helping them balance the grid and plan energy storage or backup sources. The README suggests this was built as a university project (IFT6759) where teams competed to build the best prediction model. The project is set up to be reproducible and maintainable. It includes automated code quality checks, linters that enforce consistent formatting and catch common errors, so that team members' code stays clean and readable. Training happens through a simple command-line script where you can adjust settings like learning rate, number of training cycles, and the random seed for reproducibility. Once trained, the same script lets you test the model on held-out data and submit results for evaluation against a hidden test set.