Single-feature and multi-feature, auto-regressive LSTM RNN vs linear regression for higly-volatile time series.
Project given for the first HackDay at IFIC (University of Valencia/CSIC), research institute where I completed my PhD in Physics.
A small town in Spain is sourced by solar energy thanks to a solar plant. Over a year (2015), the plant's energy output and the town's energy consumption data were gathered; and along with it, data from a nearby meteorological station.
We then come with several models that predict the amount of energy output/consumption given a set of weather conditions. In the notebook, we employ several machine learning techniques to tackle these tasks, among which we construct Long Short-Term Memory Recurrent Neural Networks.
We try and predict with an hourly basis, then a 12-hour basis and finally with a 24-hour basis, and compare. Analizing the differences has important consequences because a model will be more or less accurate depending on the time window to predict.
-
Data Preparation
- Data cleaning
- Data merging
-
Predicting energy consumption
- Data preparation
- Multi-feature linear regression
- Single-feature, auto-regressive linear regression
- Single-feature, auto-regressive long short-term memory (LSTM) recurrent neural network (RNN)
- Multi-feature, auto-regressive LSTM RNN.
-
Summary
-
References