Click the above button to launch this repository as a notebook in your browser.
Last updated of the README: 03-01-2019
This repository is currently being created, it is not yet finished.
The notebooks in the repository look the best when using Jupyter.

Multivariate Time Series Models in Keras

Introduction

This repository contains a throughout explanation on how to create different deep learning models in Keras for multivariate (tabular) time-series prediction. The data being used in this repository is from the KB-74 OPSCHALER project. The goal of this project is to do gas consumption prediction of houses on an hourly resolution, for the minor Applied Data Science at The Hague University of Applied Sciences.

Jargon

The jargon used in this repository.

• Dwelling: An individual house
• EDA: Exploratory Data Annalysis
• MVLR: Multivariate Linear Regression
• DNN: Deep Neural Network
• CNN: Convolutional Neural Network
• RNN: Recurrent Neural Network
• LSTM: Long Short-Term Memory
• GRU: Gated Recurrent Unit

The data

Available data

The original data is as follows.

Smart meter data

Parameter	Unit	Sample rate	Description
Timestamp	-	10 s	Timestamp of data telegram (set by smart meter) in local time
eMeter	kWh	10 s	Meter reading electricity delivered to client, normal tariff
eMeterReturn	kWh	10 s	Meter reading electricity delivered by client, normal tariff
eMeterLow	kWh	10 s	Meter reading electricity delivered to client, low tariff
eMeterLowReturn	kWh	10 s	Meter reading electricity delivered by client, low tariff
ePower	kWh	10 s	Actual electricity power delivered to client
ePowerReturn	kWh	10 s	Actual electricity power delivered by client
gasTimestamp	-	1 h	Timestamp of the gasMeter reading (set by smart meter) in local time
gasMeter	m3	1 h	Last hourly value (temperature converted0, gas delivered to client

Weather data

This is weather data from the KNMI weather station in Rotterdam with a sample rate of 15 minutes.
A representative from OPSCHALER says that this weather station is the most nearby most of the dwellings, the exact dwelling locations however are unknown.
The dwelling furthest away from the weather station is 103 km north east.
When this weather data is used to make predictions on the validation and test dataset (which is future data for the model), this weather data is assumed to be the 'predictions' for the weather at given timestamp.
In reality the weather predictions made by climate models should be used.

Parameter	Unit	Description
DD	degrees	Wind direction
DR	s	Precipitation time
FX	m/s	Maximum gust of wind at 10 m
FF	m/s	Windspeed at 10 m
N	okta	Cloud coverage
P	hPa	Outside pressure
Q	W/m2	Global radiation
RG	mm/h	Rain intensity
SQ	m	Sunshine duration (in minutes)
T	deg C	Temperature at 1,5 m (1 minute mean)
T10	deg C	Minimum temperature at 10 cm
TD	deg C	Dew point temperature
U	%	Relative humidity at 1,5 m
VV	m	Horizontal sight
WW	-	Weather- and station-code

Used data

The original data has been resampled to an hour, this is the data available in this repository.

Features:

• Electrical power consumption (ePower)
• Wind speed (FF)
• Rain intensity (RG)
• Temperature (T)
• Timestamp YYYY:MM:DD HH:MM:SS (datetime)

Target:

• Gas consumption (gasPower)

About the notebooks

The notebooks are written in order.
Due to this reason certain information that has been put in notebook 1 might for example not appear in notebook 2 and so on.
The hyperas MODEL.py files contains the Python scripts that use Hyperas for the hyperparameter optimazation.
The MODEL.py files contains the .py versions of the notebooks, these train quicker than training from within Jupyter (e.g. 50 epochs/s instead of 2 epochs/s, for DNN).

Notebooks in order

• 1. EDA & Feature engineering
• 2. MVLR
• 3. DNN
• 4.1 CNN & RNN Data Preperation
• 4.2 LSTM
• 4.3 GRU
• 5.1 Multivariate time-series to images
• 5.2 CNN
• 6. TimeDistributed(CNN)+RNN+DNN