This is the github repository for our research paper on "Transfer Learning for Day-Ahead Load Forecasting: A Case Study on European National Electricity Demand Time Series." (see more details here). This repository serves as a comprehensive resource for the code and experiments conducted as part of our study.
This project is implemented in MLFlow to handle the different stages in the pipeline. Each stage can be run independently as an entry point, and its inputs and outputs are stored in its respected MLflow run file -- see MLProject file for details regarding the inputs of each entry point.
| Entrypoint | Filename |
|---|---|
| main | main.py |
| load | load_raw_data.py |
| etl | etl.py |
| optuna | forecasting_model_optuna.py |
| model | forecasting_model.py |
| ensemble | forecasting_model_ensemble.py |
| eval | forecasting_model_eval.py |
| snaive* | forecasting_model_naive.py |
*snaive is an independent entrypoint used for comparative evaluation of our scenarios with naive forecasts
model_utils.py is a python file containing general-purpose functions used in more than one pipeline stages
Use the package manager pip to install MLFlow.
pip install mlflowBy executing an entrypoint, MLProject checks package dependencies (see python_env.yaml) and proceeds to install them.
Our pipeline is capable of processing multiple files from different or same countries. Data must:
- all be in a single directory provided at user input (dir_in)
- be in csv format
- contain one datetime column (named "Start") and one column containing energy load data (named "Load")
- be in 1-hour interval
- be named after the country name (e.g Greece) or code (e.g GR) they represent
As mentioned, MLProject offers a wide range of parameters that should be tuned, with regards to each stage in pipeline. While each entrypoint has its own parameters, main entrypoint contains all parameters required for any entrypoint, and distributes them accordingly:
| Parameters | Type | Default Value | Description |
|---|---|---|---|
| stages | str | 'all | comma seperated entry point names to execute from pipeline |
| stages | str | 'all' | comma-seperated containing entrypoint names to be run |
| dir_in | str | '../original_data/' | Folder path containing csv files used by the model |
| local_tz | bool | False | flag if you want local (True) or UTC (False) timezone |
| src_countries | str | 'Portugal' | csv names from dir_in used by the source model |
| tgt_countries | str | 'Portugal' | csv names from dir_in used by the target model |
| seed | str | '42' | seed used to set random state to the model |
| train_years | str | '2015,2016,2017,2018,2019' | list of years to use for training set |
| val_years | str | '2020' | list of years to use for validation set |
| test_years | str | '2021' | list of years to use for testing set |
| n_trials | str | '2' | number of trials - different tuning oh hyperparams |
| max_epochs | str | '3' | range of number of epochs used by the model |
| n_layers | str | '1' | range of number of layers used by the model |
| layer_sizes | str | "100" | range of size of each layer used by the model |
| l_window | str | '240' | range of lookback window (input layer size) used by the model |
| f_horizon | str | '24' | range of forecast horizon (output layer size) used by the model |
| l_rate | str | '0.0001' | range of learning rate used by the model |
| activation | str | 'ReLU' | activation functions experimented on by the model |
| optimizer_name | str | 'Adam' | optimizers experimented on by the model |
| batch_size | str | '1024' | batch sizes experimented on by the model |
| transfer_mode | str | "0" | indicator to use transfer learning techniques |
| num_workers | str | '2' | accelerator (cpu/gpu) processesors and threads used |
| tl_model_uri | str | None | uri path for accessing model used for transfer learning |
| n_estimators | str | '3' | number of estimators (models) used in ensembling |
| test_case | str | '1' | indicator of scenario that is being used |
Example: locally train a model in the Greek-Spanish dataset, apply AbO warm-start transfer learning on Italy and store in an experiment with name "full_pipeline":
mlflow run . --env-manager=local -P stages='all'
-P src_countries='Greece,Spain' -P tgt_countries='Italy'
-P test_case=2 -P transfer_mode=1
--experiment-name=full_pipelinetransfer_mode and test_case are integers determined by the following Enums:
class TestCase(IntEnum):
NAIVE = 0
BASELINE = 1
AbO = 2
CbO = 3
class Transfer(IntEnum):
NO_TRANSFER = 0
WARM_START = 1 The execution of a single entrypoint can be done using the "-e" flag. Example: execute "optuna" entrypoint and store run in an experiment with name "optuna_entrypoint":
mlflow run . --env-manager=local -e optuna --experiment-name=optuna_entrypointPull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.