SynthDataGen

Usage

This app is designed to be use in the following way:

1. Load the data either from ESIOS or from a local CSV file which contains a pandas.DataFrame-like table.
2. Apply the needed adjustments by
   • Applying a multiplication factor by year to all the rows
   • Upsampling or downsampling the data by means of different techniques
3. Get samples from the DataFrame, following the available probability distributions, for each row.

Workflow and usage

1. Depending on the used loader (ESIOSLoader, LocalDFLoader, etc.), the corresponding attributes are expected to be specified in the corresponding nested dictionary in the input parameters file:

   • "ESIOS": the fields for the access token, the particular indicator and the granularity for the data to be requested.
   • "LocalDF": the directory and name of the CSV file containing the DataFrame to be loaded, the variable name to get (indicator), whether to skip the first column or not, and the datetime format of the index column.

2. The Loader.getDataFromSource(...) method receive a number of parameters, which are used as filters against the loaded DataFrame. So, the resulting DataFrame will start from an

   • initial year, and consider
   • from an initial datetime (default value: 'now')
   • a number of hours ahead.
   • Besides, whether to discard the February 29 or not should also be specified.

3. The adjustments by year method receives a dictionary <year,adjustmentValue> = <int,int|float>. It is used for inflation or similar adjustments of a DataFrame. It is specified in percentage, so a 10 indicate a positive adjustment of a 10%, a -32.0 represents a negative adjustment of 32%, and a 347.89 represents just that.

4. In case a posterior upsampling or downsampling of the data wanted to be performed, the corresponding methods are used to specify the granularity when running the Adjustments.upsample(...) and Adjustments.downsample(...) methods.

   • The granularity (here frequency) should be an integer followed by a unit ('D': daily, 'H': hourly, 'T': minutely, 'S': secondly). E.g. "2T" == and entry for every 2 minutes.
   • The interpolation method and the aggregation function for upsampling and downsampling respectively, should be specified too.

5. Finally, for sampling the current data by means of te Sampling.getSamples(...) method, we should provide

   • a number of desired samples to be generated
   • and the probability distribution to consider.

Examples

A similar example has been included and extended in the ./notebooks/fullExample.ipynb Jupyter notebook.

import synthDataGen.controller as controller
from datetime import datetime

controller = controller.LocalDFLoader("./synthDataGen/settings/inputParams.json")
df = controller.getDataFromSource(initialYear=2007, datetime.now(), hoursAhead=10, include29February=False)

# DataFrame adjustments
from synthDataGen.adjustments import FactorByYear, ChangeResolution

df = FactorByYear.run(df, adjustmentsDict={2022: 10, 2021: 10, 2020: 10, 2019: 10, 2018: 10, 2017: 10, 2016: 10, 2015: 10, 2014: 10, 2013: 10, 2012: 10, 2011: 10, 2010: 10, 2009: 10, 2008: 10, 2007: 10})

# Up/down sampling
df = ChangeResolution.upsample(df, frequency="15T", method="polynomial", order=2)
df = ChangeResolution.downsample(df, frequency="2H", aggregationFunc="mean")

# Samples generation
from synthDataGen.utils import Sampling

df = Sampling.getSamples(df, 5000, "truncnorm")

Acknowledgements

© Copyright 2023, Germán Navarro $^\dagger$, Santiago Fernández Prieto $^{\ddagger,1}$, Ricardo Enríquez Miranda $^\ddagger$, Javier Hernanz Zájara $^\ddagger$

$^\dagger$ Barcelona Supercomputing Center
$^\ddagger$ Repsol Quantum Advisory Team

$^1$ Repsol-BSC Research Center

Developed within the framework of the project CUCO. Financed by the CDTI and with the support of the Spanish Ministry of Science and Innovation under the Recovery, Transformation and Resilience Plan (RTRP).