cross_val_tower_plots.py

#!/usr/bin/env python
# ~*~ coding: utf8 ~*~
from __future__ import division, print_function, unicode_literals

import datetime
import itertools

import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import pint
import scipy.optimize
import statsmodels.formula.api as smf
import xarray
from statsmodels.stats.api import anova_lm

import flux_correlation_function_fits
from correlation_function_fits import (
    CorrelationPart,
    PartForm,
    get_full_parameter_list,
    is_valid_combination,
)
from correlation_utils import get_autocorrelation_stats

CORRELATION_PARTS_LIST = [
    (day_part, dm_part, ann_part)
    for day_part, dm_part, ann_part in itertools.product(PartForm, PartForm, PartForm)
    if is_valid_combination(day_part, dm_part, ann_part)
]

# Time constants
HOURS_PER_DAY = 24
DAYS_PER_WEEK = 7
DAYS_PER_FORTNIGHT = 14
DAYS_PER_YEAR = 365.2425
DAYS_PER_DECADE = 10 * DAYS_PER_YEAR
HOURS_PER_YEAR = HOURS_PER_DAY * DAYS_PER_YEAR

N_YEARS_DATA = 4
REQUIRED_DATA_FRAC = 0.8

UREG = pint.UnitRegistry()

AMERIFLUX_MINUS_CASA_DATA = xarray.open_dataset(
    "ameriflux-and-casa-matching-data.nc4",
    chunks={"site": 30},
)

DATA_COUNTS = AMERIFLUX_MINUS_CASA_DATA["flux_difference"].count("time").load()

AUTOCORRELATION_DATA = xarray.open_dataset(
    "ameriflux-minus-casa-autocorrelation-data-all-towers.nc4",
)

CROSS_TOWER_FIT_ERROR_DS = xarray.open_dataset(
    "ameriflux-minus-casa-autocorrelation-function-fits.nc4",
)

############################################################
# Define sort orders for plots
SORT_KEYS = {
    "alphabetical": lambda site: site,
    "vegetation": lambda site: (
        AMERIFLUX_MINUS_CASA_DATA["Vegetation"].sel(site=site).values[()],
        site,
    ),
    "climate class": lambda site: (
        AMERIFLUX_MINUS_CASA_DATA["Climate_Cl"].sel(site=site).values[()],
        site,
    ),
    "latitude": lambda site: (
        AMERIFLUX_MINUS_CASA_DATA["Latitude"].sel(site=site).values[()],
        site,
    ),
    "longitude": lambda site: (
        AMERIFLUX_MINUS_CASA_DATA["Longitude"].sel(site=site).values[()],
        site,
    ),
    "mean temp": lambda site: (
        AMERIFLUX_MINUS_CASA_DATA["Mean_Temp"].sel(site=site).values[()],
        site,
    ),
    "mean precip": lambda site: (
        AMERIFLUX_MINUS_CASA_DATA["Mean_Preci"].sel(site=site).values[()],
        site,
    ),
}

############################################################
# define normalizations for the plots
NORMALIZATIONS = {
    "one": 1,
    "num pairs": (
        AUTOCORRELATION_DATA["flux_error_n_pairs"]
        .sum("time_lag")
        .rename(site="validation_tower")
    ),
    "num differences": DATA_COUNTS.rename(site="validation_tower").load(),
    "mean": CROSS_TOWER_FIT_ERROR_DS["cross_validation_error"].mean(
        ("correlation_function", "training_tower")
    ),
}


############################################################
# Create the plots for each sort order and normalization
for norm_name, norm_val in NORMALIZATIONS.items():
    normalized = CROSS_TOWER_FIT_ERROR_DS["cross_validation_error"] / norm_val
    min_err = max(
        normalized.where(
            normalized.coords["training_tower"] != normalized.coords["validation_tower"]
        ).min(),
        0,
    )
    max_err = normalized.where(
        normalized.coords["training_tower"] != normalized.coords["validation_tower"]
    ).quantile(0.95)
    for sort_name, sort_key in SORT_KEYS.items():
        sorted_towers = sorted(normalized.coords["training_tower"].values, key=sort_key)
        fig, axes = plt.subplots(9, 6, sharex=True, sharey=True, figsize=(7.5, 9))
        for ax in axes.flat:
            ax.set_visible(False)
        for corr_fun, ax in zip(normalized.coords["correlation_function"], axes.flat):
            ax.set_visible(True)
            image = ax.imshow(
                normalized.sel(
                    correlation_function=corr_fun,
                    training_tower=sorted_towers,
                    validation_tower=sorted_towers,
                ),
                vmin=min_err,
                vmax=max_err,
            )
            ax.set_title(corr_fun.coords["correlation_function_short_name"].values)
        fig.subplots_adjust(hspace=0.4)
        cbar = fig.colorbar(
            image, ax=axes[-1, -2:], orientation="horizontal", extend="both", fraction=1
        )
        cbar.set_label(
            "Cross-Validation Error{normalization:s}\n"
            "Sort order: {name:s}".format(
                name=sort_name,
                normalization=(
                    " over {0:s}".format(norm_name) if norm_name != "one" else ""
                ),
            )
        )
        fig.savefig(
            "tower-cross-validation-mismatch-over-{normalization:s}-sort-{name:s}.png".format(
                name=sort_name.replace(" ", "-"),
                normalization=norm_name.replace(" ", "-"),
            )
        )
        plt.close(fig)
        out_of_sample = (
            normalized.where(
                normalized.coords["training_tower"]
                != normalized.coords["validation_tower"]
            )
            .to_dataframe()
            .dropna()
        )
        for_regression = out_of_sample.replace(
            {
                "3-term cosine series": "Cos",
                "Exponential sine-squared": "Per",
                "Geostatistical": "Geo",
            }
        ).rename(
            dict(
                daily_cycle="day",
                annual_cycle="ann",
                annual_modulation_of_daily_cycle="day_mod",
            ),
            axis=1,
        )
        # formula = "cross_validation_error ~ has_daily_cycle + has_annual_cycle + (daily_cycle + annual_cycle + annual_modulation_of_daily_cycle) ** 2"
        # model = smf.ols(formula,  out_of_sample)
        formula0 = "cross_validation_error ~ (has_daily_cycle + has_annual_cycle) ** 2"
        model0 = smf.ols(formula0, for_regression)
        result0 = model0.fit()
        formula1 = "cross_validation_error ~ (day + day_mod + ann) ** 1"
        model1 = smf.ols(formula1, for_regression)
        result1 = model1.fit()
        formula2 = "cross_validation_error ~ (day + day_mod + ann) ** 2"
        model2 = smf.ols(formula2, for_regression)
        result2 = model2.fit()
        print("Regression against the parts of the correlation models")
        print(result1.summary())
        print("ANOVA table against has_XXX_cycle\n", anova_lm(result0, result1))