Global horizontal irradiance forecasting code and data

Published: 08-04-2021| Version 7 | DOI: 10.17632/dcj3yvzybr.7
Jorge Segarra-Tamarit,
Emilio Perez,
Hector Beltran,
Javier Perez


For reproducibility purposes, this dataset contains the data and code described in "A deep learning model for intra-day forecasting of solar irradiance using satellite-based estimations in the vicinity of a PV power plant" ( There are four main folders in the project: code, data, models and logdir. Data This folder contains all the data used from the studied location: Loc.1 (latitude=40.4º, longitude=6.0º). Sorted by year, month and day, there are three kinds of data: • The files named as just a number are 151x151 irradiance estimates matrices centered in the same location obtained from The spatial resolution is 0.03º for both latitude and longitude. • The files named Real_ are the irradiance measurements at the location • The files named CopernicusClear_ are the clear sky estimates from the CAMS McClear model Each file contains the 96 15-minute samples for the same day in Matlab format and UTC time. Code All the python scripts used to train the neural networks and perform the forecasts. The main files are: • tf1.yml: List of the modules and versions used. A clean Anaconda environment created from this file can run all the code in the project. • The script to train a new model. Changing the “paper_model_name” variable selects the kind of model to fit. • Loads a trained model and performs the forecast. Notice that the model must match the one used to train the model stored in the “training_path” folder Models This folder contains all the trained models and their forecasting results. There is also a training folder to contain the last trained model. Logdir This folder stores Tensorboard files during training How to train and test a model A new model can be trained using “”. This script has two parameters • paper_model_name: This sets the inputs to match the ones used in the models from the article. • training_path: The folder to save the trained model Then the “” script allows performing the forecasts. It is important to set the same parameters as in the “” script. This program will generate the predictions and save them in the model folder. It also plots some days, which can be modified at the bottom of the script. For instance, in order to train the TOA & all real model we would run: "python TOAallreal training" This will train the neural network and save the results in the folder models/training. After this, we would generate the results and plot some days using: “python TOAallreal training” This will save the forecasts and real values in the training folder and show figures with 1 to 6 hour forecasts The models used for the article can also be evaluated by using and targeting their folders. For instance, to evaluate the TOA & all real model used in the article, this command must be used: “python TOAallreal RtoaAllReal”