Time Series Foundation Models for Residential Load Profile Forecasting

This project aims to compare the zero-shot forecast performance of Time Series Foundation Models (TSFMs) on various scales of residential load profile (RLP) forecasting problems.

This is the official repository of paper: Link to paper.

Logging Result Procedure

Example: exp/chronos/chronos_predictor.py

1. import utility.configuration (as cf) and exp.eva_metrics (as evm) modules.
2. generate a unique exp_id for each run. exp_id=cf.generate_time_id()
3. for each sub-run (e.g. with different country, resolution.)
   1. log data configuration data_config=cf.DataConfig(country='nl',...)
   2. log model configuration model_config=cf.ModelConfig(model='chronos',...)
   3. log evaluation results eval_metrics=evm.EvaluationMetrics(...)
   4. integrate into exp_config=cf.ExperimentConfig(exp_id=exp_id, data=data_config, model=model_config, eval_metrics=eval_metrics)
   5. save to .csv exp_config.append_csv(f'result/{exp_id}.csv')

Structure

File Structure

• dataset: contains class definition of datasets used in the project.
  • XXX.py: class definition of dataset XXX.
  • ...: data preprocessing.
• model: contains the (wrapper) class of the TSFMs used in the project.
  • YYY.py: class definition of TSFM YYY.
• utility: contains utility functions used in the project, including data
  • argument_parser.py: argument parser for the project.
  • configuration.py: configuration class definition.
• configs: .yaml configuration files.

Configuration Usage

The configuration module defines a basic configuration class that can be extended to contain configuration settings for data, model, etc. The base class allows for easy conversion between dictionary, configuration object, and .yaml file.

Configuration Hierachy

• ExperimentConfig
  • general experiment-specific settings such as exp_id.
  • data: DataConfig. configuration for data.
  • model: ModelConfig. configuration for model.
  • ...