09_timeseries.rst

{{ header }}

.. ipython:: python

    import pandas as pd
    import matplotlib.pyplot as plt

Data used for this tutorial:

• Air quality data

  For this tutorial, air quality data about NO_2 and Particulate matter less than 2.5 micrometers is used, made available by OpenAQ and downloaded using the py-openaq package. The air_quality_no2_long.csv" data set provides NO_2 values for the measurement stations FR04014, BETR801 and London Westminster in respectively Paris, Antwerp and London.

  To raw data

  .. ipython:: python
  
      air_quality = pd.read_csv("data/air_quality_no2_long.csv")
      air_quality = air_quality.rename(columns={"date.utc": "datetime"})
      air_quality.head()
  
  

  .. ipython:: python
  
      air_quality.city.unique()
  
  

How to handle time series data with ease

Using pandas datetime properties

• I want to work with the dates in the column datetime as datetime objects instead of plain text

  .. ipython:: python
  
      air_quality["datetime"] = pd.to_datetime(air_quality["datetime"])
      air_quality["datetime"]
  
  

  Initially, the values in datetime are character strings and do not provide any datetime operations (e.g. extract the year, day of the week,…). By applying the to_datetime function, pandas interprets the strings and convert these to datetime (i.e. datetime64[ns, UTC]) objects. In pandas we call these datetime objects similar to datetime.datetime from the standard library as :class:`pandas.Timestamp`.

Note

As many data sets do contain datetime information in one of the columns, pandas input function like :func:`pandas.read_csv` and :func:`pandas.read_json` can do the transformation to dates when reading the data using the parse_dates parameter with a list of the columns to read as Timestamp:

pd.read_csv("../data/air_quality_no2_long.csv", parse_dates=["datetime"])

Why are these :class:`pandas.Timestamp` objects useful? Let’s illustrate the added value with some example cases.

What is the start and end date of the time series data set we are working with?

.. ipython:: python

    air_quality["datetime"].min(), air_quality["datetime"].max()

Using :class:`pandas.Timestamp` for datetimes enables us to calculate with date information and make them comparable. Hence, we can use this to get the length of our time series:

.. ipython:: python

    air_quality["datetime"].max() - air_quality["datetime"].min()

The result is a :class:`pandas.Timedelta` object, similar to datetime.timedelta from the standard Python library and defining a time duration.

To user guide

The various time concepts supported by pandas are explained in the user guide section on :ref:`time related concepts <timeseries.overview>`.

• I want to add a new column to the DataFrame containing only the month of the measurement

  .. ipython:: python
  
      air_quality["month"] = air_quality["datetime"].dt.month
      air_quality.head()
  
  

  By using Timestamp objects for dates, a lot of time-related properties are provided by pandas. For example the month, but also year, quarter,… All of these properties are accessible by the dt accessor.

To user guide

An overview of the existing date properties is given in the :ref:`time and date components overview table <timeseries.components>`. More details about the dt accessor to return datetime like properties are explained in a dedicated section on the :ref:`dt accessor <basics.dt_accessors>`.

• What is the average NO_2 concentration for each day of the week for each of the measurement locations?

  .. ipython:: python
  
      air_quality.groupby(
          [air_quality["datetime"].dt.weekday, "location"])["value"].mean()
  
  

  Remember the split-apply-combine pattern provided by groupby from the :ref:`tutorial on statistics calculation <10min_tut_06_stats>`? Here, we want to calculate a given statistic (e.g. mean NO_2) for each weekday and for each measurement location. To group on weekdays, we use the datetime property weekday (with Monday=0 and Sunday=6) of pandas Timestamp, which is also accessible by the dt accessor. The grouping on both locations and weekdays can be done to split the calculation of the mean on each of these combinations.

  !DANGER!

  As we are working with a very short time series in these examples, the analysis does not provide a long-term representative result!

• Plot the typical NO_2 pattern during the day of our time series of all stations together. In other words, what is the average value for each hour of the day?

  .. ipython:: python
  
      fig, axs = plt.subplots(figsize=(12, 4))
      air_quality.groupby(air_quality["datetime"].dt.hour)["value"].mean().plot(
          kind='bar', rot=0, ax=axs
      )
      plt.xlabel("Hour of the day");  # custom x label using Matplotlib
      @savefig 09_bar_chart.png
      plt.ylabel("$NO_2 (µg/m^3)$");
  
  

  Similar to the previous case, we want to calculate a given statistic (e.g. mean NO_2) for each hour of the day and we can use the split-apply-combine approach again. For this case, we use the datetime property hour of pandas Timestamp, which is also accessible by the dt accessor.

Datetime as index

In the :ref:`tutorial on reshaping <10min_tut_07_reshape>`, :meth:`~pandas.pivot` was introduced to reshape the data table with each of the measurements locations as a separate column:

.. ipython:: python

    no_2 = air_quality.pivot(index="datetime", columns="location", values="value")
    no_2.head()

Note

By pivoting the data, the datetime information became the index of the table. In general, setting a column as an index can be achieved by the set_index function.

Working with a datetime index (i.e. DatetimeIndex) provides powerful functionalities. For example, we do not need the dt accessor to get the time series properties, but have these properties available on the index directly:

.. ipython:: python

    no_2.index.year, no_2.index.weekday

Some other advantages are the convenient subsetting of time period or the adapted time scale on plots. Let’s apply this on our data.

• Create a plot of the NO_2 values in the different stations from the 20th of May till the end of 21st of May

  .. ipython:: python
      :okwarning:
  
      @savefig 09_time_section.png
      no_2["2019-05-20":"2019-05-21"].plot();
  
  

  By providing a string that parses to a datetime, a specific subset of the data can be selected on a DatetimeIndex.

To user guide

More information on the DatetimeIndex and the slicing by using strings is provided in the section on :ref:`time series indexing <timeseries.datetimeindex>`.

Resample a time series to another frequency

• Aggregate the current hourly time series values to the monthly maximum value in each of the stations.

  .. ipython:: python
  
      monthly_max = no_2.resample("ME").max()
      monthly_max
  
  

  A very powerful method on time series data with a datetime index, is the ability to :meth:`~Series.resample` time series to another frequency (e.g., converting secondly data into 5-minutely data).

The :meth:`~Series.resample` method is similar to a groupby operation:

• it provides a time-based grouping, by using a string (e.g. M, 5H,…) that defines the target frequency
• it requires an aggregation function such as mean, max,…

To user guide

An overview of the aliases used to define time series frequencies is given in the :ref:`offset aliases overview table <timeseries.offset_aliases>`.

When defined, the frequency of the time series is provided by the freq attribute:

.. ipython:: python

    monthly_max.index.freq

• Make a plot of the daily mean NO_2 value in each of the stations.

  .. ipython:: python
      :okwarning:
  
      @savefig 09_resample_mean.png
      no_2.resample("D").mean().plot(style="-o", figsize=(10, 5));
  
  

To user guide

More details on the power of time series resampling is provided in the user guide section on :ref:`resampling <timeseries.resampling>`.

REMEMBER

• Valid date strings can be converted to datetime objects using to_datetime function or as part of read functions.
• Datetime objects in pandas support calculations, logical operations and convenient date-related properties using the dt accessor.
• A DatetimeIndex contains these date-related properties and supports convenient slicing.
• Resample is a powerful method to change the frequency of a time series.

To user guide

A full overview on time series is given on the pages on :ref:`time series and date functionality <timeseries>`.