This is a very simple implementation of utilizing large language models (e.g., Llama-3.2-1B-Instruct) for time-series forecasting in the scenarios of electric vehicle charging. If it is helpful to your research, please cite our papers:
Haohao Qu, Han Li, Linlin You, Rui Zhu, Jinyue Yan, Paolo Santi, Carlo Ratti, Chau Yuen. (2024) ChatEV: Predicting electric vehicle charging demand as natural language processing. Transportation Research Part D: Transport and Environment. Paper in TRD
For simple implementaton, we need five major packages, namely torch, pandas, numpy, transformers, and argparse. You can install these useful wheels by:
pip install -r requirements.txtTo get access to Meta-Llama models, we need to apply a "hf_token" key through https://huggingface.co/settings/tokens
Then replace input a correct "hf_token" in Line 99 of the "utils.py" file.
hf_token = "Your_HF_TOKEN"ps: Or you can download a local model through https://www.llama.com/llama-downloads/
To conduct a simple implementation (inference only), we can run the "simple.py" file.
python simple.pyTo conduct a simple finetuning implementation, we can run the "finetune.py" file.
python finetune.pyAll code for a complete implementation of ChatEV (including finetuning, validation, and testing) is included in the "code" folder. Besides the five packages for simple version, more environments are required for the full implementation: [argparse, lightning, scikit-learn].
pip install argparse, lightning, scikit-learnPlease remember to change your path to the "code" folder.
cd codeAlso replace input a correct "hf_token" in Line 228 of the "model_interface.py" file
hf_token = "Your_HF_TOKEN"We can run the "main.py" file to fintune a Llama model for EV charging data prediction:
python main.py- If you wanna load a checkpoint for finetuning, you can do so:
python main.py --ckpt --ckpt_name='last'- If you wanna load a checkpoint for testing, you can do so:
python main.py --ckpt --ckpt_name='last' --test_only- Train the model in a few-shot scienario:
python main.py --few_shot --few_shot_ratio=0.2- Train the model using a simple and effective meta-learning approach, First-Order Reptile:
python main.py --meta_learningMore configurations can be found in the "parse.py" file.
- If you uncounter a problem of slow downloading, you can set a mirror source at the command terminal:
export HF_ENDPOINT=https://hf-mirror.com