TimeSeriesClustering is a Julia implementation of unsupervised learning methods for time series datasets. It provides functionality for clustering and aggregating, detecting motifs, and quantifying similarity between time series datasets. The software provides a type system for temporal data, and provides an implementation of the most commonly used clustering methods and extreme value selection methods for temporal data. It provides simple integration of multi-dimensional time-series data (e.g. multiple attributes such as wind availability, solar availability, and electricity demand) in a single aggregation process. The software is applicable to general time series datasets and lends itself well to a multitude of application areas within the field of time series data mining.

The TimeSeriesClustering package was originally developed to perform time series aggregation for energy systems optimization problems. By reducing the number of time steps used in the optimization model, using representative periods leads to significant reductions in computational complexity of these problems. The packages was previously known as ClustForOpt.jl.

The package has three main purposes:

1. Provide a simple process of finding representative periods (reducing the number of observations) for time-series input data, with implementations of the most commonly used clustering methods and extreme value selection methods.
2. Provide an interface between representative period data and application (e.g. optimization problem) by having representative period data stored in a generalized type system.
3. Provide a generalized import feature for time series, where variable names, attributes, and node names are automatically stored and can then be used later when the reduced time series is used in the application at hand (e.g. in the definition of sets of the optimization problem).

In the domain of energy systems optimization, an example problem that uses TimeSeriesClustering for its input data is the package CapacityExpansion, which implements a scalable generation and transmission capacity expansion problem.

The TimeSeriesClustering package follows the clustering framework presented in Teichgraeber and Brandt, 2019. The package is actively developed, and new features are continuously added. For a reproducible version of the methods and data of the original paper by Teichgraeber and Brandt, 2019, please refer to v0.1 (including shape based methods such as k-shape and dynamic time warping barycenter averaging).

This package is developed by Holger Teichgraeber @holgerteichgraeber and Elias Kuepper @YoungFaithful.

This package runs under julia v1.0 and higher. Install using:

import Pkg
Pkg.add("TimeSeriesClustering")

Documentation (Stable): Please refer to this documentation for details on how to use TimeSeriesClustering the current version of TimeSeriesClustering. This is the documentation of the default version of the package.

Documentation (Development): If you like to try the development version of TimeSeriesClustering, please refer to this documentation.

See NEWS for significant breaking changes when updating from one version of TimeSeriesClustering to another.

If you find TimeSeriesClustering useful in your work, we kindly request that you cite the following paper (link):

  @article{Teichgraeber2019,
  author = {Holger Teichgraeber and Adam Brandt},
  title = {Clustering methods to find representative periods for the optimization of energy systems: An initial framework and comparison},
  journal = {Applied Energy},
  volume = {239},
  pages = {1283–1293},
  year = {2019},
  doi = {https://doi.org/10.1016/j.apenergy.2019.02.012},
  }

This quick start guide introduces the main concepts of using TimeSeriesClustering. For more detail on the different functionalities that TimeSeriesClustering provides, please refer to the subsequent chapters of the documentation or the examples in the examples folder, specifically workflow_introduction.jl.

Generally, the workflow consists of three steps:

• load data
• find representative periods (clustering + extreme period selection)
• optimization

After TimeSeriesClustering is installed, you can use it by saying:

using TimeSeriesClustering

The first step is to load the data. The following example loads hourly wind, solar, and demand data for Germany (1 region) for one year.

ts_input_data = load_timeseries_data(:CEP_GER1)

The output ts_input_data is a ClustData data struct that contains the data and additional information about the data.

ts_input_data.data # a dictionary with the data.
ts_input_data.data["wind-germany"] # the wind data (choose solar, el_demand as other options in this example)
ts_input_data.K # number of periods

The second step is to cluster the data into representative periods. Here, we use k-means clustering and get 5 representative periods.

clust_res = run_clust(ts_input_data;method="kmeans",n_clust=5)
ts_clust_data = clust_res.clust_data

The ts_clust_data is a ClustData data struct, this time with clustered data (i.e. less representative periods).

ts_clust_data.data # the clustered data
ts_clust_data.data["wind-germany"] # the wind data. Note the dimensions compared to ts_input_data
ts_clust_data.K # number of periods

If this package is used in the domain of energy systems optimization, the clustered input data can be used as input to an optimization problem. The optimization problem formulated in the package CapacityExpansion can be used with the data clustered in this example.