This is the Github repository for the code used in the paper Deep Learning for Predicting Microbial Fuel Cell Energy Output. In this paper, we explore ways to improve the viability of intermittenly powered computing systems, with a focus on Soil Microbial Fuel Cells (SMFCs). SMFCs are an energy source which generate power from natural microbial interactions within soil, and they are a promising energy source for many types of environmental sensing tasks, such as smart farming and wildfire prevention. Very little work currently exists that attempts to model the relationship between soil conitions and SMFC energy generation, and this work is the first to use machine learning to do so.
All relevant datasets and models are stored in Hugging Face at https://huggingface.co/datasets/adunlop621/Soil_MFC/tree/main
In order for the code to run properly, it is neccesary to download the datasets from Hugging Face and place them in the same directory as the code with the following file names. Dataset 1, collected from a deployment at Stanford University, is stored in stanfordMFCDataset.zip, which expands into the directory rocket4. Dataset 2, collected from a deployment at UC Santa Cruz, is stored in the ucscMFCDataset directory. In order to load the pretrained models used in the paper, it is neccesary to download the trained_models directory.
If you want to just load and evaluate pretrained models instead of training from scratch, use the code in Pretrained.ipynb.
We define two main types of models in this work, based off the datasets used to train them: Type 1 models are trained on Dataset 1, and Type 2 models are trained on Dataset 2.
Type1.ipynb, Type2.ipynb, and SNN.ipynb contain the code used to train all all the models in the paper, as well as load pretrained models and run time series rolling validation.
Pretrained models have the following naming conventions: model type (Type 1, 2, 1A, 1B, etc.), time horizon (3min, 5min, etc.), and quantile (quant5, quant50, etc.). For example, a Type 1A model predicting lower bound values in time intervals of 15 minutes would be called type1A_15min_quant5.
The same naming convention holds for SNN models, with the model type simply being snn, ie. snn_15min_quant50.