Modeling with Machine Learning
Aug. - Dec. 2024
Two projects done in a class covering machine learning tools for modeling energy transport and conversion processes. Final project was self-directed to explore time-series data prediction.
-
Machine Learning
-
Modeling
-
Thermodynamics
What I did:
-
Project 1: Photovoltaic (PV) Cells
-
Training neural network model to predict the voltage-current relationship for a PV cell based on operating characteristics
-
Training a second model to predict which configuration in a 4-PV-cell system maximizes power output
-
Validating both models to achieve a mean absolute error below 0.03
-

Training and test loss as function of epoch when implementing a learning rate scheduler in Keras

Model-predicted power outputs vs actual PV cell outputs for training dataset, validation dataset, and a dataset outside the range of provided data (Higflux)

Surface plot of the final neural network model predictions for power output as a function of load resistance, R, and incident direct normal solar radiation intensity, ID
-
Project 2: Physics Inspired Neural Network (PINN) for 2D Heat Conduction
-
Predicting a temperature field for a 2D material domain using machine learning and a custom loss function
-
Incorporating the custom loss function taking into account thermodynamic boundary conditions, min/max temperatures, and Laplacian interior temperature rules
-
Training the model to predict the temperature field with a loss of 0.009
-

Design of custom loss function incorporating boundary conditions, min/max temperature values, and the heat transport equation. (Van P. Carey)

Illustration of the material being modeled with boundary conditions and positions of temperature readings in black. (Van P. Carey)

Model-predicted temperature distribution in the material domain

-
Project 3: Running Speed Estimation from Leg-Work IMU
-
Designing a neural network to predict a runner's speed based on acceleration data from an inertial mass unit (IMU)
-
Explored design parameters for the model including RNN, LSTM, masking, and dropout layers
-
Determined the optimal model configuration to obtain a validation mean absolute error of 4.69
-

Running acceleration IMU data my teammate collected to be used in training the model. (Douglas Hutchings)

Neural network design parameters for the best performing model

Predictions of the best performing model vs actual treadmill's (and runner's) speed