2025 / UC Berkeley (ME249)
Solar PV Modeling at High Flux
Extended-range power prediction with neural networks
Overview
Extended the PV performance model to flux levels up to 1850 W/m² and evaluated alternate network depths.
Problem
Predict solar panel voltage and power across higher irradiance levels while maintaining generalization.
Role
Course project (ME249)
Timeline
Dec 2025
Tools
Python / Keras
Tags
Regression / Energy / ML
Data
- - Combined datasets from Project 3 and new high-flux measurements
- - Inputs: air temperature, irradiance, load resistance
Approach
- - Merged datasets and normalized by median values
- - Trained a baseline 3-layer network and a deeper 4-layer variant
- - Generated power surface plots vs load and irradiance
Evaluation
- - MAE targets defined in coursework (<0.025)
- - Log-log plots for train/validation fit checks
Results
- - Surface plots used to compare model behavior across flux ranges
- - Model depth trade-offs documented in report
Deployment & Monitoring
- - Notebook-based workflow for reproducible plotting
Limitations
- - Evaluation focused on course-provided data only
Gallery
Illustrative prediction fit (synthetic)
Illustrative power surface (synthetic)
Repro Steps
- - See CodeP4.* notebooks and ME249Project4F25.pdf
Next Steps
- - Add uncertainty estimates for high-flux extrapolation
- - Test alternative regularization to reduce overfit risk