Merge pull request #171 from sadielbartholomew/plotting-improvements

Improve plot output (hatching, unit formatting, mean lines, etc.)

Author: andreww

cats/plotting.py

@@ -1,10 +1,13 @@
 try:
     import matplotlib.pyplot as plt
     import matplotlib.dates as mdates
+    from matplotlib.patches import Rectangle
+    from matplotlib.ticker import FuncFormatter
+
     have_matplotlib = True
 except ImportError:
     have_matplotlib = False
-    
+
 
 def plotplan(CI_forecast, output):
     """
@@ -13,55 +16,238 @@ def plotplan(CI_forecast, output):
     if not have_matplotlib:
         print("To plot graphs you must import matplotlib")
         print("e.g. \"pip install 'climate-aware-task-scheduler[plots]'\"")
-        return None
-    
+        return
+
     # Just for now pull the CI forecast apart... probably belongs as method...
     values = []
     times = []
     now_values = []
     now_times = []
     opt_values = []
     opt_times = []
+
     # For our now and optimal series, start with the starting data (interpolated)
     opt_times.append(output.carbonIntensityOptimal.start)
     opt_values.append(output.carbonIntensityOptimal.start_value)
     now_times.append(output.carbonIntensityNow.start)
     now_values.append(output.carbonIntensityNow.start_value)
+
     # Build the three time series of point from the API
     for point in CI_forecast:
         values.append(point.value)
         times.append(point.datetime)
-        if (point.datetime >= output.carbonIntensityOptimal.start and
-            point.datetime <= output.carbonIntensityOptimal.end):
+        if (
+            point.datetime >= output.carbonIntensityOptimal.start
+            and point.datetime <= output.carbonIntensityOptimal.end
+        ):
             opt_values.append(point.value)
             opt_times.append(point.datetime)
-        if (point.datetime >= output.carbonIntensityNow.start and
-            point.datetime <= output.carbonIntensityNow.end):
+        if (
+            point.datetime >= output.carbonIntensityNow.start
+            and point.datetime <= output.carbonIntensityNow.end
+        ):
             now_values.append(point.value)
             now_times.append(point.datetime)
+
     # For our now and optimal series, end with the end data (interpolated)
     opt_times.append(output.carbonIntensityOptimal.end)
     opt_values.append(output.carbonIntensityOptimal.end_value)
     now_times.append(output.carbonIntensityNow.end)
     now_values.append(output.carbonIntensityNow.end_value)
 
+    # Determine if there is any window overlap, since then we add an extra
+    # element to the legend to make clear how the overlap region looks but
+    # otherwise we can leave it out. There will be an overlap if the end of
+    # the 'now' time is more than the start of the optimal time since the
+    # optimal is only ever time shifted forwards.
+    windows_overlap = (
+        output.carbonIntensityNow.end > output.carbonIntensityOptimal.start
+    )
+
     # Make the plot (should probably take fig and ax as opt args...)
-    fig, ax = plt.subplots()
-    ax.fill_between(times, 0.0, values, alpha=0.2, color='b')
-    ax.fill_between(now_times, 0.0, now_values, alpha=0.6,  color='r')
-    ax.fill_between(opt_times, 0.0, opt_values, alpha=0.6, color='g')
-
-    ax.text(0.125, 1.075, f"Mean carbon intensity if job started now: {output.carbonIntensityNow.value:.2f} gCO2eq/kWh",
-             transform=ax.transAxes, color='red')
-    ax.text(0.125, 1.025, f"Mean carbon intensity at optimal time: {output.carbonIntensityOptimal.value:.2f} gCO2eq/kWh",
-             transform=ax.transAxes, color='green')
-
-    ax.set_xlabel("Time (dd-mm-yy hh)")
-    ax.xaxis.set_major_formatter(mdates.DateFormatter("%d-%m-%y %H:%M"))
-    ax.xaxis.set_minor_formatter(mdates.DateFormatter("%d-%m-%y %H:%M"))
-    ax.set_ylabel("Forecast carbon intensity (gCO2eq/kWh)")
-    ax.grid(True)
+
+    # Use an accessibility-approved colour scheme to ensure plot is
+    # comprehendable for all (though also use hatching and clear text so
+    # that no information is only encoded by colour choice/view)
+    plt.style.use("tableau-colorblind10")
+    forecast_colour = "tab:blue"
+    now_colour = "tab:red"
+    optimal_colour = "tab:green"
+
+    # Supply a figsize a bit different o the default (6.4, 4.8) to make the
+    # plot a little bit more square else the x-axis label gets a bit cut off
+    # due to long datetime labels
+    fig, ax = plt.subplots(figsize=(7.0, 5.5))
+
+    # Filling under curves for the forecast, run now time and optimal run time
+    ax.fill_between(
+        times,
+        0.0,
+        values,
+        alpha=0.2,
+        color=forecast_colour,
+        edgecolor=forecast_colour,
+        label="Forecast",
+    )
+    # Show 'now' window in red with black hatch lines for contrast
+    ax.fill_between(
+        now_times,
+        0.0,
+        now_values,
+        alpha=0.6,
+        color=now_colour,
+        label="If job started now",
+        hatch="///",
+        edgecolor="k",
+    )
+    # Show 'optimal' window in green with hatch lines in opposite direction
+    # but also in black: for any overlapping regions on the two windows, this
+    # therefore results in a distinguishable cross-hatch pattern
+    ax.fill_between(
+        opt_times,
+        0.0,
+        opt_values,
+        alpha=0.6,
+        color=optimal_colour,
+        label="Optimal job window",
+        hatch="\\\\\\",
+        edgecolor="k",
+    )
+
+    # In case the 'now' and 'optimal' windows overlap, is nice to show just
+    # how that looks on legend, namely crosshatch in an (ugly) khaki colour
+    # that represents the mixture of the transparent red and green. To do
+    # this, use matplotlib's patches to create a proxy object i.e. patch.
+    if windows_overlap:
+        overlap_patch = Rectangle(
+            # Arbitrary huge number to ensure dummy patch is outside plot area
+            (1e10, 1e10),
+            1,
+            1,
+            facecolor="#6f8d4a",  # mix colour from image colour picker tool
+            edgecolor="k",
+            hatch="////\\\\\\\\",
+            label="Overlap area (now + optimal)",
+            transform=ax.transAxes,  # << use axes coords instead of data coords
+        )
+        ax.add_patch(overlap_patch)
+        handles, labels = ax.get_legend_handles_labels()
+        handles.append(overlap_patch)
+        labels.append(overlap_patch.get_label())
+        ax.legend(handles=handles, labels=labels)
+
+    now_value = output.carbonIntensityNow.value
+    optimal_value = output.carbonIntensityOptimal.value
+    # To avoid having to check if a user has (La)TeX available, to format the
+    # units, use matploltib built-in lightweight TeX parser 'Mathtext' via
+    # '$' symbols. Allows subscript on 2, etc. Needs a raw string to work.
+    units = r"$\mathrm{g\,CO_{2}\,eq\;kWh^{-1}}$"
+
+    ax.text(
+        0.5,
+        1.05,
+        f"Projected carbon intensity ({units}) mean...",
+        ha="center",
+        va="bottom",
+        fontsize=14,
+        transform=ax.transAxes,
+    )
+    ax.text(
+        0.45,
+        1.0,
+        f"...if job started now: {now_value:.2f}",
+        ha="right",
+        va="bottom",
+        color=now_colour,
+        fontsize=14,
+        transform=ax.transAxes,
+    )
+    # Separator to divide the two described figures ('now' and 'optimal')
+    ax.text(
+        0.5,
+        1.0,
+        r"$\to$",
+        ha="center",
+        va="bottom",
+        color="black",
+        fontsize=14,
+        transform=ax.transAxes,
+    )
+    ax.text(
+        0.55,
+        1.0,
+        f"...at optimal time: {optimal_value:.2f}",
+        ha="left",
+        va="bottom",
+        color=optimal_colour,
+        fontsize=14,
+        transform=ax.transAxes,
+    )
+
+    # For a nice illustration of CI saved, plot the lines corresponding to
+    # the mean value for the 'now' and 'optimal' cases:
+    plt.axhline(
+        y=now_value,
+        color=now_colour,
+        linestyle="--",
+        alpha=0.4,
+        label="Mean if started now",
+    )
+    plt.axhline(
+        y=optimal_value,
+        color=optimal_colour,
+        linestyle="--",
+        alpha=0.4,
+        label="Mean for optimal window",
+    )
+
+    # Include subtle markers at each data point, in case it helps to
+    # distinguish forecast points from the trend (esp. useful if there)
+    # is a similar trend across/for 1 hour or more i.e. 3+ data points
+    ax.scatter(times, values, color=forecast_colour, s=8, alpha=0.3)
+    ax.scatter(now_times, now_values, color=now_colour, s=8, alpha=0.3)
+    ax.scatter(opt_times, opt_values, color=optimal_colour, s=8, alpha=0.3)
+
+    def tick_formatting(x, pos):
+        """Format datetimes so the x-axis labels become more readable.
+
+        Namely, only show the full 'yy-mm-dd hh:mm' format date at the start
+        of a day i.e. at hh:mm = 00:00, doing so in bold to provide emphasis.
+        Otherwise show an ellipsis and then the hh:mm component only. This
+        makes it much quicker for a reader to parse the date and time
+        axis span and partitioning.
+        """
+        dt = mdates.num2date(x)
+
+        # Would otherwise always full date at the first tick, as well as
+        # at every point we get to a new day i.e. 00:00 (every four major
+        # ticks), but then the date runs off the figure area to the left. So
+        # it should be clear enough with two full dates plotted - which will
+        # be the case always since we have a 48 hour window
+        if dt.hour == 0 and dt.minute == 0:
+            # Note \text{} required around '-' symbol else it is
+            # interpreted as a minus sign and becomes so long with spacing
+            # around that it pushes the x axis label off the figure below
+            date_bold = dt.strftime(r"\mathbf{%y\text{-}%m\text{-}%d}")
+            time_part = dt.strftime("%H:%M")
+            return rf"${date_bold}\ {time_part}$"
+        else:
+            # All other ticks get ellipsis + time
+            time_part = dt.strftime("%H:%M")
+            return rf"$\ldots\ {time_part}$"
+
+    # The x axis label needs some padding at the figure foot else it gets a
+    # bit cut off due to the length of some datetime x labels
+    ax.set_xlabel(r"Time ($\mathbf{yy\text{-}mm\text{-}dd}$ hh:mm)")
+    ax.xaxis.set_major_formatter(FuncFormatter(tick_formatting))
+    ax.set_ylabel(rf"Forecast carbon intensity ({units})")
     ax.label_outer()
+
+    ax.grid(True)
+    ax.legend()
+
     fig.autofmt_xdate()
+    ax.set_ylim(bottom=0)  # start y-axis at 0, negative CI not possible!
+
+    plt.subplots_adjust(bottom=0.20)
     plt.show()
-    return None