generate_training_data.py

"""
Author: Andreas Bott
year: 2024
mail: andreas.bott@eins.tu-darmstadt.de or andi-bott@web.de

This script generates random samples of power plant and demand inputs and corresponding grid states. The samples are
stored in csv files which are numbered consecutively.

csv file format:
    columns: [Q1, Q2, Q3, Q4, Q0], [T1, T2, T3, T4, T0], [[grid state vector == (T, mf, p, T_end)]]

In order to generate samples faster, an importance-sampling-approach is used which will be published in the paper:
    "Efficient Training of Learning-Based Thermal Power Flow for 4th Generation District Heating Grids"
    by Andreas Bott, Mario Beykirch, and Florian Steinke

    currently under review at the journal "Energy", preprint available at: https://arxiv.org/abs/2403.11877

The script is designed to be run in parallel, where each instance computes a fraction of the samples.
:param n_samples: number of samples generated by all instances
:param n_instances: number of instances running in parallel
:param instant: integer argument to distinguish the different instances expected to be passed as sys.argv[1]
"""

import sys
import lib_dnn_soc.DNN_lib as NN_lib
import lib_dnn_soc.utility as util
import time


def __main__(results_file, grid_settings, n_samples, save_file, verbose=True):
    Results = util.ResFile(name=results_file, path='./results/', print=verbose)
    Results.log(f'Run DNN training script for {grid_identifier}, current time: {time.ctime()}')
    SE, d_prior_dist, T_prior_dist, cycles, grid = util.setup_training(grid_identifier, grid_settings)
    # setup importance sampler, i.e. compute proxy distribution
    IS = NN_lib.ImportanceSampler(d_prior_dist, T_prior_dist, SE, cycles, grid, results_file=save_file)
    IS.setup()
    # generate training samples
    IS.generate_training_samples(n_samples, include_slack=True, file_spec=save_file, verbose=verbose)
    Results.log(f'Finished generating {n_samples} training data for {grid_identifier}, current time: {time.ctime()}')


if __name__ == '__main__':
    grid_identifier = 'ladder5'
    results_file = 'results.out'
    from Settings import grid_settings, DNN_settings

    n_instances = 1  # number of simultaneous runs, each computing a fraction of the samples
    n_samples = DNN_settings['n_training'] + DNN_settings['n_val'] + DNN_settings['n_test']
    n_per_file = 2000

    # in order to avoid overwriting existing files, n_samples / n_per_file * n_runs should be an integer
    assert n_samples % (n_per_file * n_instances) == 0, (f'n_samples / (n_per_file * n_runs) should be an integer; '
                                                         f'n_samples: {n_samples}, n_per_file: {n_per_file}, '
                                                         f'n_runs: {n_instances}')

    # number of samples computed by this instance:
    n_samples = int(n_samples / n_instances)
    instant = 0 if len(sys.argv) < 2 else int(sys.argv[1])


    # choose save file dynamically to fill up files subsequently. Starting file depends on the instance number.
    def save_file(idx):
        zerofile = instant * (n_samples // n_per_file)
        folder = f'DNN_model/data/{grid_identifier}_{grid_settings["d_prior_type"]}/'
        file = f'{grid_identifier}_{idx // n_per_file + zerofile}.csv'
        return folder + file


    __main__(results_file, grid_settings, n_samples, save_file, verbose=True)