Multivariate Forecasting of Energy Production

Introduction

Forecasting models are specialized tools designed to predict future values based on historical data of a given variable. They find application in diverse fields such as stock market trends, energy production, and energy demand analysis.

In the context of renewable energy sources, where energy generation is weather-dependent, a continuous and reliable energy supply cannot be guaranteed. This project specifically focuses on forecasting energy production in a solar farm, aiming to comprehend the energy flux and make informed decisions to prevent blackouts.

This endeavor involved the creation of various predictive models, employing multivariate approaches with tools like Prophet, as well as utilizing deep learning techniques such as LSTM, GRU, and Convolutional Networks.

Data information:

The data used come from the following Kaggle data set.
The variables available were wind speed, sunshine, air pressure, radiation, air temperature, relative air humidity, and the system production.

Data treatment

Data was loaded and analyzed to determine the most suitable time period for forecasting.
The dataset was resampled based on the identified time period.
Multivariate and univariate forecasting models (Prophet, LSTM, GRU, and Convolutional) were developed for prediction purposes.
The training data comprised 80% of the initial dataset, while the remaining 20% was reserved for testing the models.
The relevance of features for multivariate forecasting was assessed using multiple Prophet models with various feature combinations.
The most impactful features were incorporated into the multivariate forecasting models, which were then evaluated using the mean absolute error.
A comprehensive overview of this modeling process is provided in the accompanying flowchart.

Results

The results shows that Multivariate forecast performs better than the univariate forecasting in all models.

LSTM models gave almost the same performance for univariate and multivariate forecast.

Recurrent Neural Networks are not the most capable model to make the forecasting, in both scenarios, the Convolutional Neural Networks got better results.

The results are shown in the following picture.

Public code is available in the following GitHub repo.

Follow me on my social media channels to know more about my projects.

Get In Touch

Pujilí, Cotopaxi, Ecuador

sebitas.alejo@hotmail.com

Quick Links

Home About Us Our Services Contact Us