def testfeq_equal(self):
"""feq should return true when they are equal"""
val1 = 1.1234
val2 = 1.1235
self.assertTrue(tb.feq(val1, val2, 0.0001))
numpy.testing.assert_array_equal(
[False, True, False, False],
tb.feq([1., 2., 3., 4.],
[1.25, 2.05, 2.2, 500.1],
0.1)
)
def addmodelerror_old(dd, A, y, L):
"""
this routine will add a standard error to the ensemble states
"""
Lgrid = dd['model']['Lgrid']
dL = Lgrid[1] - Lgrid[0]
nens = int(dd['kalman']['nens'])
#print(y, L)
radius = 1
for Lcenter, yval in zip(L, y):
L1 = np.max((Lcenter - radius, Lgrid[0]))
L2 = np.min((Lcenter + radius, Lgrid[-1]))
#print Lcenter, Lcenter+radius, Lgrid[-1], min((Lcenter+radius,Lgrid[-1])), (L2-L1)/dL
NLs = int(round((L2 - L1) / dL) + 1)
for Lpos in np.linspace(L1, L2, NLs):
#print dL, L1, L2, NLs
index = np.where(feq(Lpos, Lgrid)) # use float point comparison
stdev = 0.5 * np.abs(yval - np.mean(A[index, :]))
#print Lpos
center = np.reshape(A[index, :], (nens))
A[index, :] = np.random.normal(center, scale=stdev)
# now check if any below 1e-99 and repeat
for i in range(nens):
icnt = 0
while A[index, i] < 1e-99:
#A[index,i] = np.random.normal( center[i], scale=stdev)
A[index, i] = 10**np.random.normal(np.log10(center[i]),
scale=0.001)
icnt += 1
if icnt > 1000: print('too many iterations')
return A
def addmodelerror_old2(dd, A, y, L):
"""
this routine will add a standard error to the ensemble states
"""
from numpy import where, random, log10, mean, reshape, abs, max, arange
from numpy import linspace, min
Lgrid = dd['model']['Lgrid']
dL = Lgrid[1] - Lgrid[0]
nens = int(dd['kalman']['nens'])
#print y, L
radius = 1
for Lcenter, yval in zip(L, y):
L1 = max((Lcenter - radius, Lgrid[0]))
L2 = min((Lcenter + radius, Lgrid[-1]))
#print Lcenter, Lcenter+radius, Lgrid[-1], min((Lcenter+radius,Lgrid[-1])), (L2-L1)/dL
NLs = int(round((L2 - L1) / dL) + 1)
for Lpos in linspace(L1, L2, NLs):
#print dL, L1, L2, NLs
index = where(feq(Lpos, Lgrid)) # use float point comparison
stdev = 0.5 * abs(yval - mean(A[index, :]))
#print Lpos
center = reshape(A[index, :], (nens))
A[index, :] = random.normal(center, scale=stdev)
# now check if any below 1e-99 and repeat
for i in range(nens):
icnt = 0
while A[index, i] < 1e-99:
A[index, i] = random.normal(center[i], scale=stdev)
icnt += 1
if icnt > 1000: print('too many iterations')
return A
def average_window(PSDdata, Lgrid):
"""
combine observations on same L shell in
Parameters
==========
model :
PSDdata :
HAp :
Returns
=======
out :
Examples
========
"""
# sort observations first in L
idx = PSDdata['Lstar'].argsort()
Lobs = PSDdata['Lstar'][idx]
y = PSDdata['PSD'][idx]
# map Lobs onto closest grid point
for i, obs in enumerate(y):
Lobs[i] = Lgrid[np.argmin(np.abs(Lobs[i] - Lgrid))]
# identify unique grid-points of Lobs
tmpLobs = np.unique(Lobs)
# declare average observation array
tmpy = np.zeros_like(tmpLobs)
# run through all unique grid-points and compute average observation
for i, iL in enumerate(tmpLobs):
# identify idex of each unique grid-point
idx = np.where(feq(iL, Lobs))
# compute average observation for each unique grid-point
tmpy[i] = np.average(y[idx])
# walk through all grid points and find how many obs available
#for i, iL in enumerate(Lgrid):
# idx = np.where( iL == Lobs[:])[0]
# assign Lobs and y
Lobs = tmpLobs
y = tmpy
return Lobs, y
def fill_gaps(data,
fillval=9999999,
sigma=5,
winsor=0.05,
noise=False,
constrain=False):
'''Fill gaps in input data series, using interpolation plus noise
The noise approach is based on Owens et al. (Space Weather, 2014).
data - input numpy ndarray-like
fillval - value marking fill in the time series
sigma - width of gaussian filter for finding fluctuation CDF
winsor - winsorization threshold, values above p=1-winsor and below p=winsor are capped
noise - Boolean, if True add noise to interpolated region, if False use linear interp only
constrain - Boolean, if True
'''
# identify sequences of fill in data series
gaps = np.zeros((len(data), 2), dtype=int)
k = 0
for i in range(1, len(data) - 1):
# Single space gap/fillval
if (tb.feq(data[i], fillval)) and (~tb.feq(data[i + 1], fillval)) and (
~tb.feq(data[i - 1], fillval)):
gaps[k][0] = i
gaps[k][1] = i
k += 1
# Start of multispace gap/fillval
elif (tb.feq(data[i], fillval)) and (~tb.feq(data[i - 1], fillval)):
gaps[k][0] = i
# End of multispace gap/fillval
elif (tb.feq(data[i], fillval)) and (~tb.feq(data[i + 1], fillval)):
gaps[k][1] = i
k += 1
gaps = gaps[:k]
#if no gaps detected
if k == 0:
return data
# fill gaps with linear interpolation
for gap in gaps:
a = data[gap[0] - 1]
b = data[gap[1] + 1]
dx = (b - a) / (gap[1] - gap[0] + 2)
for i in range(gap[1] - gap[0] + 1):
data[gap[0] + i] = a + dx * (i + 1)
if noise:
# generate CDF from delta var
series = data.copy()
smooth = gaussian_filter(series, sigma)
dx = series - smooth
dx.sort()
p = np.linspace(0, 1, len(dx))
# "Winsorize" - all delta-Var above/below threshold at capped at threshold
dx[:p.searchsorted(0. + winsor)] = dx[p.searchsorted(0. + winsor) + 1]
dx[p.searchsorted(1. - winsor):] = dx[p.searchsorted(1. - winsor) - 1]
# draw fluctuations from CDF and apply to linearly filled gaps
for gap in gaps:
for i in range(gap[1] - gap[0] + 1):
data[gap[0] + i] += dx[p.searchsorted(random.random())]
# cap variable if it should be strictly positive (e.g. number density)
# use lowest measured value as floor
if constrain and series.min() > 0.0:
data[data < series.min()] = series.min()
return data
def testfeq_notequal(self):
"""feq should return false when they are not equal"""
val1 = 1.1234
val2 = 1.1235
self.assertFalse(tb.feq(val1, val2, 0.000005))
def assimilate(self, method='EnKF', inflation=0):
"""
Assimilates data for the radiation belt model using the Ensemble
Kalman Filter. The algorithm used is the SVD method presented by
Evensen in 2003 (Evensen, G., Ocean dynamics, 53, pp.343--367, 2003).
To compensate for model errors, three inflation algorithms are
implemented. The inflation methodology is specified by the
'inflation' argument, and the options are the following:
inflation == 0: Add model error (perturbation for the ensemble)
around model state values only where observations are available
(DEFAULT).
inflation == 1: Add model error (perturbation for the ensemble)
around observation values only where observations are available.
inflation == 2: Inflate around ensemble average for EnKF.
Prior to assimilation, a set of data values has to be speficied by
setting the start and end dates, and time step, using the setup_ticks
funcion of the radiation belt model:
>>> import spacepy
>>> import datetime
>>> from spacepy import radbelt
>>> start = datetime.datetime(2002,10,23)
>>> end = datetime.datetime(2002,11,4)
>>> delta = datetime.timedelta(hours=0.5)
>>> rmod.setup_ticks(start, end, delta, dtype='UTC')
Once the dates and time step are specified, the data is added using the
add_PSD function:
>>> rmod.add_PSD()
The observations are averaged over the time windows, whose interval is
give by the time step.
Once the dates and data are set, the assimiation is performed using the
'assimilate' function:
>>> rmod.assimilate(inflation=1)
This function will add the PSDa values, which are the analysis state of
the radiation belt using the observations within the dates. To plot the
analysis simply use the plot funtion:
>>> rmod.plot(values=rmod.PSDa,clims=[-10,-6],Lmax=False,Kp=False,Dst=False)
"""
import spacepy.data_assimilation
import spacepy.sandbox.PSDdata as PD
import copy as c
# add PSD observations with add_PSD,
# this has to be done to the class
# module when running the RBmodel.
# add_PSD will add the PSDdata to the class
# which is a dictionary for the data that has been added
# debugging command
#pdb.set_trace()
# setup method
assert method in [
'EnKF', 'insert'
], 'data assimilation method=' + method + ' not implemented'
nTAI = len(self.ticks)
# enKF method
if method == 'EnKF':
da = spacepy.data_assimilation.ensemble()
# initialize A with initial condition
# the initial condition is ones, have to change this to initialize
# the ensemble with perturbed reference state
A = np.ones((self.NL, da.Nens)) * self.PSDinit[:, np.newaxis]
self.PSDf = np.zeros((self.PSDinit.shape[0], nTAI))
self.PSDa = np.zeros((self.PSDinit.shape[0], nTAI))
self.PSDa[:, 0] = self.PSDinit
self.PSDf[:, 0] = self.PSDinit
# diagnostic tools:
# observations-minus-background
self.PSD_omb = [''] * (nTAI)
# observations-minus-analysis
self.PSD_oma = [''] * (nTAI)
# analysis-minus-background
self.PSD_amb = np.zeros((self.PSDinit.shape[0], nTAI))
# add model error (perturbation) in the ensemble initial conditionp.
#A = da.add_model_error(self, A, self.PSDdata[0])
std = 0.35
normal = np.random.randn(da.Nens)
for iens in np.arange(da.Nens):
A[:, iens] = A[:, iens] + std * normal[iens] * A[:, iens]
A[np.where(A < self.MIN_PSD)] = self.MIN_PSD
# ==========================================
# DEBUG
#np.savetxt('ensemble_IC.dat',A)
#np.savetxt('model_IC.dat',self.PSDinit)
#np.savetxt('model_grid_IC.dat',self.Lgrid)
# ==========================================
# create temporary RB class instance
rbtemp = c.copy(self)
# time loop
for i, Tnow, Tfut in zip(
np.arange(nTAI - 1) + 1, self.ticks[:-1], self.ticks[1:]):
# make forcast and add model error
# make forecast using all ensembles in A
iens = 0
for f in A.T:
rbtemp.ticks = st.Ticktock([Tnow.UTC[0], Tfut.UTC[0]],
'UTC')
rbtemp.PSDinit = f.copy()
rbtemp.evolve()
#rbtemp.PSDdata = self.PSDdata[i-1]
A[:, iens] = rbtemp.PSD[:, 1].copy()
iens += 1
# save result in ff
Tnow = Tfut
self.PSDf[:, i] = np.mean(A, axis=1)
# verify that there are data points within the interval, if
# there are data points then extract average observations
# within the window, if not return empty observation array y
# and Lobs.
if len(self.PSDdata[i - 1]) > 0:
# get observations for time window ]Tnow-Twindow,Tnow]
Lobs, y = spacepy.data_assimilation.average_window(
self.PSDdata[i - 1], self.Lgrid)
else:
y = np.array([])
Lobs = np.array([])
print(Lobs)
print(y)
# ==========================================
# DEBUG
#np.savetxt('obs_location_IC.dat',Lobs)
#np.savetxt('obs_IC.dat',y)
#pdb.set_trace()
# ==========================================
# then assimilate otherwise do another forcast
if len(y) > 0:
### check for minimum PSD values
### A[np.where(A<self.MIN_PSD)] = self.MIN_PSD
### # insert observations directly
### A = da.add_model_error_obs(self, A, Lobs, y)
### # dictionary
### self.PSDa[:,i] = np.mean(A, axis=1)
# INFLATION SCHEMES
if inflation == 0:
print(
'inflation around model state values at observation locations'
)
# Add model error (perturbation for the ensemble) around model
# state values. This acts as an inflation scheme for EnKF
A = da.add_model_error(self, A, self.PSDdata[i - 1])
elif inflation == 1:
print(
'inflation around observation values at observation locations'
)
# Add model error (perturbation for the ensemble) around
# observation values. This acts as an inflation scheme for EnKF
A = da.add_model_error_obs(self, A, Lobs, y)
elif inflation == 2:
print('inflation around ensemble average')
# Inflate around ensemble average for EnKF
# ensemble average
ens_avg = np.mean(A, axis=1)
# inflation factor
inflation_factor = 1.8
# loop over ensemble members and inflate
iens = 0
for ens in A.T:
# inflate ensemble
A[:, iens] = inflation_factor * (ens -
ens_avg) + ens_avg
iens += 1
A[np.where(A < self.MIN_PSD)] = self.MIN_PSD
# prepare assimilation analysis
# project ensemble states to obs. grid
HA = da.getHA(self, Lobs, A)
# measurement perturbations ensemble
Psi = da.getperturb(self, y)
# ensemble of innovation vectors
Inn = da.getInnovation(y, Psi, HA)
# calculate ensemble perturbation HA' = HA-HA_mean
HAp = da.getHAprime(HA)
# calculate prior diagnostics
# observation minus background
omb = y - np.average(HA, axis=1)
self.PSD_omb[i] = {
'Lobs': Lobs,
'y': y,
'omb': omb,
}
xf_avg = np.average(A, axis=1)
# now call the main analysis routine
if len(y) == 1:
A = da.EnKF_oneobs(A, Psi, Inn, HAp)
else:
A = da.EnKF(A, Psi, Inn, HAp)
# check for minimum PSD values
A[np.where(A < self.MIN_PSD)] = self.MIN_PSD
# average A from analysis step and save in results
# dictionary
self.PSDa[:, i] = np.mean(A, axis=1)
# calculate posterior diagnostics
# observation minus analysis
Hanalysis = da.getHA(self, Lobs, A)
oma = y - np.average(Hanalysis, axis=1)
self.PSD_oma[i] = {
'Lobs': Lobs,
'y': y,
'omb': oma,
}
# analysis minus background
xa_avg = np.average(A, axis=1)
self.PSD_amb[:, i] = xa_avg - xf_avg
#pdb.set_trace()
# print assimilated result
Hx = np.zeros_like(y)
for iL, Lstar in enumerate(Lobs):
idx = np.where(tb.feq(self.Lgrid, Lstar))
Hx[iL] = self.PSDa[idx, i]
print(Hx)
print(HA)
elif len(y) == 0:
print('no observations within this window')
self.PSDa[:, i] = self.PSDf[:, i]
continue #
# print message
print('Tnow: ', self.ticks[i].ISO)
# insert obsrvations for data assimilation
elif method == 'insert':
da = spacepy.data_assimilation.ensemble()
A = np.ones((self.NL, 1)) * self.PSDinit[:, np.newaxis]
self.PSDf = np.zeros((self.PSDinit.shape[0], nTAI))
self.PSDa = np.zeros((self.PSDinit.shape[0], nTAI))
self.PSDa[:, 0] = self.PSDinit
self.PSDf[:, 0] = self.PSDinit
# add model error (perturbation) in the ensemble initial condition.
std = 0.15
normal = np.random.randn(self.NL)
A[:, 0] = (1.0 + std * normal) * A[:, 0]
A[np.where(A < self.MIN_PSD)] = self.MIN_PSD
rbtemp = c.copy(self)
# time loop
for i, Tnow, Tfut in zip(
np.arange(nTAI - 1) + 1, self.ticks[:-1], self.ticks[1:]):
# evolve solution
rbtemp.ticks = st.Ticktock([Tnow.UTC[0], Tfut.UTC[0]], 'UTC')
rbtemp.PSDinit = A[:, 0]
rbtemp.evolve()
A[:, 0] = rbtemp.PSD[:, 1]
# save result in ff
Tnow = Tfut
self.PSDf[:, i] = A[:, 0]
# verify that there are data points within the interval, if
# there are data points then extract average observations
# within the window, if not return empty observation array y
# and Lobs.
if len(self.PSDdata[i - 1]) > 0:
Lobs, y = spacepy.data_assimilation.average_window(
self.PSDdata[i - 1], self.Lgrid)
else:
y = np.array([])
Lobs = np.array([])
print(Lobs)
print(y)
#pdb.set_trace()
# then assimilate otherwise do another forcast
if len(y) > 0:
# check for minimum PSD values
A[np.where(A < self.MIN_PSD)] = self.MIN_PSD
# create index of obs location
idx = np.array([], dtype=int)
for Lval in Lobs:
Lidx = np.where(fp_equality.eq(Lval, self.Lgrid))[0]
idx = np.append(idx, np.array([int(Lidx)]))
# insert observations directly
A[idx, 0] = y
#A = da.add_model_error_obs(self, A, Lobs, y)
# dictionary
self.PSDa[:, i] = A[:, 0]
elif len(y) == 0:
print('no observations within this window')
self.PSDa[:, i] = self.PSDf[:, i]
continue #
# print message
print('Tnow: ', self.ticks[i].ISO)
def add_PSD_obs(self, time=None, PSD=None, Lstar=None, satlist=None):
"""
add PSD observations
Parameters
----------
time : Ticktock datetime array
array of observation times
PSD : list of numpy arrays
PSD observational data for each time. Each entry in the list is a
numpy array with the observations for the corresponding time
Lstar : list of numpy arrays
Lstar location of each PSD observations. Each entry in the list is
a numpy array with the location of the observations for the
corresponding time
satlist : list of satellite names
Returns
-------
out : list of dicts
Information of the observational data, where each entry
contains the observations and locations of observations for each
time specified in the time array. Each list entry is a dictionary
with the following information:
Ticks : Ticktock array
time of observations
Lstar : numpy array
location of observations
PSD : numpy array
PSD observation values
sat : list of strings
satellite names
MU : scalar value
Mu value for the observations
K : scalar value
K value for the observations
"""
import pdb
assert 'ticks' in self.__dict__ , \
"Provide tick range with 'setup_ticks'"
Tgrid = self.ticks
nTAI = len(Tgrid)
# initialize PSDdata list
self.PSDdata = [''] * (nTAI - 1)
if (PSD == None):
# PSD data not provided,
# extract from database
import spacepy.sandbox.PSDdata as PD
for i, Tnow, Tfut in zip(np.arange(nTAI - 1), Tgrid[:-1],
Tgrid[1:]):
start_end = spacepy.time.Ticktock([Tnow.UTC[0], Tfut.UTC[0]],
'UTC')
self.PSDdata[i] = PD.get_PSD(start_end, self.MU, self.K,
satlist)
else:
# PSD data arrays provided
# model grid
Lgrid = self.Lgrid
itime = 0
# loop over time defined for the model integration
for i, Tnow, Tfut in zip(np.arange(nTAI - 1), Tgrid[:-1],
Tgrid[1:]):
# empty array
lstar = np.array([], dtype=float)
time_idx = np.array([], dtype=int)
# loop over observation time
for itime in np.arange(len(time)):
if (Tnow <= time[itime] and time[itime] <= Tfut):
#print 'match!! '
#print itime
#print i
#print time[itime]
#print Tnow
#print Tfut
# concatenate to lstar
lstar = np.concatenate((lstar, Lstar[itime]))
lstar = np.unique(lstar)
#idx = lstar.argsort()
#pdb.set_trace()
# add time index
time_idx = np.append(time_idx, itime)
#end loop
#pdb.set_trace()
if (time_idx.shape[0] > 0):
# initialize PSD array
psd = np.zeros_like(lstar)
# initialize number of obs array
num_obs = np.zeros_like(lstar)
# sort time index
time_idx = np.unique(time_idx)
# loop over time index
for itime in time_idx:
# sort observations
idx = Lstar[itime].argsort()
tmplstar = Lstar[itime][idx]
tmppsd = PSD[itime][idx]
# run through all unique grid-points and compute
# average observation
for j, iL in enumerate(lstar):
# identify idex of grid-point
idx = np.where(tb.feq(iL, tmplstar))
# assign observation for grid-point
psd[j] = psd[j] + tmppsd[idx]
# add for number of observations
num_obs[j] = num_obs[j] + 1.0
psd = psd / num_obs
# assign time for observations
#Ticks = time[itime]
Ticks = Tfut
# determine position of observations
#lstar = Lstar[itime]
# determine observations PSD
#psd = PSD[itime]
# provide MU
MU = self.MU * np.ones_like(lstar)
# provide K
K = self.K * np.ones_like(lstar)
# empy satellite
sat = ['']
# add to dictionary
self.PSDdata[i] = {'Ticks':Ticks, 'Lstar':lstar, \
'PSD':psd, 'sat':sat, \
'MU':MU, 'K':K}
# adjust initial conditions to these PSD values
mval = np.mean(self.PSDdata[0]['PSD'])
self.PSDinit = mval * np.exp(-(self.Lgrid - 5.5)**2 / 0.8)
return
def assimilate_JK(dd):
"""
this version is currently not working
main function to assimilate all data provided in init
Parameters
==========
model :
PSDdata :
HAp :
Returns
=======
out :
Examples
========
"""
np.random.seed(123)
Lgrid = dd['model']['Lgrid']
dd['model']['initPSD'] = np.array(4e-7 * np.exp(-(Lgrid - 5)**2 / 2))
nens = dd['kalman']['nens']
NL = dd['model']['ngrid']
f0 = np.array(dd['model']['initPSD'])
Tgrid = dd['model']['Tgrid']
dd['results'] = {}
dd['results']['fa'] = {}
dd['results']['ff'] = {}
# --- initialize some stuff
# broadcast all f0 into A and randomize
A = np.ones((NL, nens)) * f0[:, np.newaxis]
#A = 10**np.random.normal(np.log10(A),0.3) # watch out for negative values
dd['results']['fa']['Tgrid'] = np.zeros(len(Tgrid))
dd['results']['ff']['Tgrid'] = np.zeros(len(Tgrid))
dd['results']['fa']['PSD'] = np.zeros((f0.shape[0], Tgrid.shape[0]))
dd['results']['fa']['PSD'][:, 0] = f0
dd['results']['fa']['Tgrid'][0] = Tgrid[0]
dd['results']['ff']['PSD'] = np.zeros((f0.shape[0], Tgrid.shape[0]))
dd['results']['ff']['PSD'][:, 0] = f0
dd['results']['ff']['Tgrid'][0] = Tgrid[1]
minPSD = np.min(A) / 100
# time loop to assimilate everything
for Tnow, tcnt in zip(Tgrid, range(len(Tgrid) - 1)):
# make forecast using all ensembles in A
icnt = 0
for f in A.T:
fnew = forecast(dd, f, Tnow)
A[:, icnt] = fnew[:]
icnt += 1
#print(np.log10(np.mean(A,axis=1)))
Tnow = Tnow + dd['model']['Twindow']
tcnt = tcnt + 1
dd['results']['ff']['PSD'][:, tcnt] = np.mean(A, axis=1)
dd['results']['ff']['Tgrid'][tcnt] = Tnow
# get observations for time window ]Tnow-Twindow,Tnow]
dd, L, y = getobs4window(dd, Tnow)
# check if len(y) ==0
if len(y) == 0:
dd['results']['fa']['PSD'][:, tcnt] = np.mean(A, axis=1)
dd['results']['fa']['Tgrid'][tcnt] = Tnow
continue
# perturb observations so that ensemble get sufficient spread
HA = getHA(dd, L,
A) # project ensemble states to obs. grid: HA(nobs,nens)
e = np.zeros((len(y), nens)) # this is Psi in Evensen 2003
D = np.zeros((len(y), nens))
err_z = 0.3
for yval, iobs in zip(y, range(len(y))):
relstd = yval * err_z
rnd = np.random.normal(yval * np.ones(nens), relstd)
rnd[np.where(rnd < minPSD)] = minPSD
e[iobs, :] = rnd - yval
D[iobs, :] = yval + e[iobs, :]
# add model error
relstd = np.zeros(nens)
for yval, iobs in zip(y, range(len(y))):
idx = np.where(feq(L[iobs], Lgrid))
relstd[:] = 0.5 * (yval - HA[iobs, :])
A[idx, :] = np.random.normal(HA[iobs, :], np.abs(relstd))
idx2 = np.where(A[idx, :] < minPSD)
A[idx, idx2] = minPSD # add min flux here
# get residual or innovation in the classical sense
Res = D - HA # dim = (nobs,nens)
# error of covariant matrix
ffmean = np.mean(A, axis=1)
V = np.zeros((NL, nens))
for iens in range(nens):
V[:, iens] = A[:, iens] - ffmean
Pfxx = np.dot(V, V.T) / float(nens)
Pfyy = np.dot(e, e.T) / float(nens)
# get Kalman denominator (H*Pfxx*HT + Pfyy)^-1 by inversion
PH, HPH = getHPH(dd, L, Pfxx)
Rinv = np.linalg.inv(HPH + Pfyy)
Adj = np.dot(np.dot(PH, Rinv), Res)
# update all ensemble members
A = A + Adj
# check for negative fluxes
idx = np.where(A < minPSD)
A[idx] = minPSD
#
# average A from analysis step and save in results dictionary
#print y, np.mean(HA,axis=1), np.mean(getHA(init,L,A),axis=1)
dd['results']['fa']['PSD'][:, tcnt] = np.mean(A, axis=1)
dd['results']['fa']['Tgrid'][tcnt] = Tnow
# print message
print('Tnow: ', rbtools.TAInum2ISOdate(Tnow))
#print np.log10(np.mean(A,axis=1))[30]
# copy some results into ['kalman'] key
dd['kalman']['fa'] = dd['results']['fa']
dd['kalman']['ff'] = dd['results']['ff']
return dd
def add_model_error_obs(self, model, A, Lobs, y):
"""
this routine will add a standard error to the ensemble states
Parameters
==========
model :
A :
Lobs :
y :
Returns
=======
out :
Examples
========
"""
# debugging command
#pdb.set_trace()
Lgrid = model.Lgrid
dL = Lgrid[1] - Lgrid[0]
nens = A.shape[1]
#nens = int(self.Nens)
#print y, L
# in units of L (should not be smaller than dL)
radius = dL * 0.9
radius = 1
# create L vector where random noise will be added
#erridx = np.array([False]*len(Lgrid))
erridx = np.array([], dtype=int)
for Lval in Lobs:
Lidx = np.where(feq(Lval, Lgrid))[0]
erridx = np.append(erridx, np.array([int(Lidx)]))
# compute residual
# calculate the rel. average uncertainty based on (obs -
# ensemble)/obs
rst = np.zeros((len(Lobs), nens))
for iens, member in enumerate(A.T):
rst[:, iens] = np.abs(y - A[erridx, iens])
# perturb with normal distribution
# around observation values
stdev = np.std(rst, axis=1)
# debug output files
#np.savetxt('Lobs.dat',Lobs)
#np.savetxt('y.dat',y)
#np.savetxt('Lgrid.dat',Lgrid)
#np.savetxt('A.dat',A)
#np.savetxt('erridx.dat',erridx)
#np.savetxt('rst.dat',rst)
#np.savetxt('stdev.dat',stdev)
pert_obs = (1.0 + 0.30 * np.random.normal(size=y.shape[0])) * y
for i, idx in enumerate(erridx):
if (stdev[i] > 0.0):
A[idx, :] = np.random.normal(y[i], stdev[i], nens)
#A[idx,:] = np.random.normal( pert_obs[i], stdev[i], nens )
print('done')
return A
def add_model_error(self, model, A, PSDdata):
"""
this routine will add a standard error to the ensemble states
Parameters
==========
model :
A :
PSDdata :
Returns
=======
out :
Examples
========
"""
Lgrid = model.Lgrid
dL = Lgrid[1] - Lgrid[0]
nens = int(self.Nens)
#print y, L
# in units of L (should not be smaller than dL)
radius = dL * 0.9
radius = 1
# create L vector where random noise will be added
# erridx = np.array([False]*len(Lgrid))
# for Lval in PSDdata['Lstar']:
# L1 = np.max((Lval-radius, Lgrid[0]))
# L2 = np.min((Lval+radius, Lgrid[-1]))
# Lidx = np.where( (Lgrid > L1) & (L2 > Lgrid) )[0]
# erridx[Lidx] = True
# calculate the rel. average uncertainty based on (obs -
# modelfcst)/obs
# fcst = np.mean(A, axis=1)
# rst = np.zeros( len(PSDdata['Lstar']) )
# for Lval, yval, i in zip(PSDdata['Lstar'], PSDdata['PSD'], range(len(PSDdata['Lstar'])) ):
# idx = np.argmin(np.abs(Lval - Lgrid))
# rst[i] = np.abs( yval - fcst[idx] )
# stdev = np.mean(rst)
# for iens in range(nens):
# f = A[erridx, iens]
# A[erridx, iens] = np.random.normal( f, scale = stdev )
##TODO: Check why above code was commented... this block was unindented for consistency, but should it be here?
# create L vector where random noise will be added
erridx = np.array([], dtype=int)
for Lval in PSDdata['Lstar']:
Lidx = np.where(feq(Lval, Lgrid))[0]
erridx = np.append(erridx, np.array([int(Lidx)]))
# calculate the rel. average uncertainty based on (obs -
# modelfcst)/obs
fcst = np.mean(A, axis=1)
rst = np.zeros((len(PSDdata['Lstar']), nens))
for iens, member in enumerate(A.T):
rst[:, iens] = np.abs(PSDdata['PSD'] - member[erridx])
stdev = 2.0 * np.std(rst, axis=1)
# debug output files
#np.savetxt('Lobs_init.dat',PSDdata['Lstar'])
#np.savetxt('y_init.dat',PSDdata['PSD'])
#np.savetxt('Lgrid_init.dat',Lgrid)
#np.savetxt('A_init.dat',A)
#np.savetxt('erridx_init.dat',erridx)
#np.savetxt('rst_init.dat',rst)
#np.savetxt('stdev_init.dat',stdev)
for i, idx in enumerate(erridx):
if (stdev[i] > 0.0):
A[idx, :] = np.random.normal(fcst[i], stdev[i], nens)
return A
