Time adjustment

Time resolution adjustment (resampling)

Models have a default timestep length (defined implicitly by the timesteps of the model's time series data). This default resolution can be adjusted by specifying time resolution adjustment in the model configuration, for example:

config:
  init:
    time_resample: 6H

In the above example, this would resample all time series data to 6-hourly timesteps. Any pandas-compatible rule describing the target resolution can be used.

Time clustering

By supplying a file linking dates in your model timeseries with representative days, it is possible to cluster your timeseries:

config:
  init:
    time_cluster: cluster_days.csv

When using representative days, you may want to enable a number of additional constraints to improve how carriers are stored between representative days, based on the study undertaken by Kotzur et al.. These constraints require a new decision variable storage_inter_cluster, which tracks storage between all the dates of the original timeseries. This particular math - detailed here - can be enabled by including storage_inter_cluster in your list of additional math.

We no longer provide the functionality to infer representative days from your timeseries. Instead, we recommend you use other timeseries processing tools applied to your input CSV data or your built model dataset (model.inputs). If you want cluster the timeseries data yourself, we recommend these tools:

• tsam: designed specifically for large-scale energy system models.
• scikit-learn: a general machine learning library that has a clustering module. We were previously using this in our internal clustering.
• tslearn: a timeseries-focussed machine learning library.

Using the tsam library to cluster your timeseries

In this example, we will find 12 representative days of a clustered timeseries and save those to file. For more configuration options, see the tsam documentation.

import tsam.timeseriesaggregation as tsam
import calliope

# Load data at full time resolution
model = calliope.Model(...)
# Get all timeseries data from model, with timesteps on the rows and all other dimensions on the columns
raw_data = (
    model.inputs[[
        k for k, v in model.inputs.data_vars.items()
        if "timesteps" in v.dims and len(v.dims) > 1
    ]]
    .to_dataframe()
    .stack()
    .unstack("timesteps")
    .T
)
aggregation = tsam.TimeSeriesAggregation(
  raw_data, noTypicalPeriods=12, hoursPerPeriod=24, clusterMethod="hierarchical"
)
typPeriods = aggregation.createTypicalPeriods()
matched_indices = aggregation.indexMatching()
representative_dates = (
    raw_data
    .resample("1D")
    .first()
    .iloc[aggregation.clusterCenterIndices]
    .index
)
cluster_days = (
    matched_indices
    .resample("1D")
    .first()
    .PeriodNum
    .apply(lambda x: representative_dates[x])
)
cluster_days.to_csv("/absolute_path/to/clusters.csv")

model_clustered = calliope.Model(..., time_cluster="/absolute_path/to/clusters.csv")

Note

Resampling and clustering can be applied together. Resampling of your timeseries will take place before clustering.

Warning

When using time clustering, the resulting timesteps will be assigned different weights depending on how long a period of time they represent. Weights are used for example to give appropriate weight to the operational costs of aggregated typical days in comparison to individual extreme days, if both exist in the same processed time series. The weighting is accessible in the model data, e.g. through model.inputs.timestep_weights.

The interpretation of results when weights are not 1 for all timesteps requires caution. Production values are not scaled according to weights, but costs are multiplied by weight, in order to weight different timesteps appropriately in the objective function. This means that costs and outflow values are not consistent without manually post-processing them by either multiplying outflows by weight (outflows would then be inconsistent with capacities) or dividing costs by weight. The computation of levelised costs and of capacity factors takes weighting into account, so these values are consistent and can be used as usual.