import numpy as np
from pyFTS.common import FLR, fts
from pyFTS.models.nonstationary import common, flrg
[docs]class ConventionalNonStationaryFLRG(flrg.NonStationaryFLRG):
"""First Order NonStationary Fuzzy Logical Relationship Group"""
def __init__(self, LHS, **kwargs):
super(ConventionalNonStationaryFLRG, self).__init__(1, **kwargs)
self.LHS = LHS
self.RHS = set()
[docs] def get_key(self):
return self.LHS
[docs] def append_rhs(self, c, **kwargs):
self.RHS.add(c)
def __str__(self):
tmp = self.LHS + " -> "
tmp2 = ""
for c in sorted(self.RHS):
if len(tmp2) > 0:
tmp2 = tmp2 + ","
tmp2 = tmp2 + c
return tmp + tmp2
[docs]class NonStationaryFTS(fts.FTS):
"""NonStationaryFTS Fuzzy Time Series"""
def __init__(self, **kwargs):
super(NonStationaryFTS, self).__init__(**kwargs)
self.name = "Non Stationary FTS"
self.shortname = "NSFTS"
self.detail = ""
self.flrgs = {}
self.method = kwargs.get('method','conditional')
self.is_high_order = False
if self.partitioner is not None:
self.append_transformation(self.partitioner.transformation)
if self.method == 'conditional':
self.min_stack = [0, 0, 0]
self.max_stack = [0, 0, 0]
self.uod_clip = False
self.order = 1
self.min_order = 1
self.max_lag = 1
self.inputs = []
self.forecasts = []
self.residuals = []
self.variance_residual = 0.
self.mean_residual = 0.
self.memory_window = kwargs.get("memory_window", 5)
[docs] def generate_flrg(self, flrs, **kwargs):
for flr in flrs:
if flr.LHS.name in self.flrgs:
self.flrgs[flr.LHS.name].append_rhs(flr.RHS.name)
else:
self.flrgs[flr.LHS.name] = ConventionalNonStationaryFLRG(flr.LHS.name)
self.flrgs[flr.LHS.name].append_rhs(flr.RHS.name)
def _smooth(self, a):
return .1 * a[0] + .3 * a[1] + .6 * a[2]
[docs] def train(self, data, **kwargs):
if self.method == 'unconditional':
window_size = kwargs.get('parameters', 1)
tmpdata = common.fuzzySeries(data, self.sets,
self.partitioner.ordered_sets,
window_size, method='fuzzy')
else:
tmpdata = common.fuzzySeries(data, self.sets,
self.partitioner.ordered_sets,
method='fuzzy', const_t=0)
flrs = FLR.generate_non_recurrent_flrs(tmpdata)
self.generate_flrg(flrs)
if self.method == 'conditional':
self.forecasts = self.forecast(data, no_update=True)
self.residuals = np.array(data[1:]) - np.array(self.forecasts[:-1])
self.variance_residual = np.var(self.residuals) # np.max(self.residuals
self.mean_residual = np.mean(self.residuals)
self.residuals = self.residuals[-self.memory_window:].tolist()
self.forecasts = self.forecasts[-self.memory_window:]
self.inputs = np.array(data[-self.memory_window:]).tolist()
[docs] def conditional_perturbation_factors(self, data, **kwargs):
npart = len(self.partitioner.sets)
_max = 0
_min = 0
if data < self.original_min:
_min = data - self.original_min if data < 0 else self.original_min - data
elif data > self.original_max:
_max = data - self.original_max if data > 0 else self.original_max - data
self.min_stack.pop(2)
self.min_stack.insert(0, _min)
_min = min(self.min_stack)
self.max_stack.pop(2)
self.max_stack.insert(0, _max)
_max = max(self.max_stack)
_range = (_max - _min)/2
translate = np.linspace(_min, _max, npart)
var = np.std(self.residuals)
var = 0 if var < 1 else var
loc = (self.mean_residual + np.mean(self.residuals))
location = [_range + w + loc + k for k in np.linspace(-var,var, npart) for w in translate]
scale = [abs(location[0] - location[2])]
scale.extend([abs(location[k - 1] - location[k + 1]) for k in np.arange(1, npart)])
scale.append(abs(location[-1] - location[-3]))
perturb = [[location[k], scale[k]] for k in np.arange(npart)]
return perturb
def _fsset_key(self, ix):
return self.partitioner.ordered_sets[ix]
def _affected_sets(self, sample, perturb):
if self.method == 'conditional':
affected_sets = [[ct, self.sets[self._fsset_key(ct)].membership(sample, perturb[ct])]
for ct in np.arange(len(self.partitioner.sets))
if self.sets[self._fsset_key(ct)].membership(sample, perturb[ct]) > 0.0]
if len(affected_sets) == 0:
if sample < self.partitioner.lower_set().get_lower(perturb[0]):
affected_sets.append([0, 1])
elif sample > self.partitioner.upper_set().get_upper(perturb[-1]):
affected_sets.append([len(self.sets) - 1, 1])
else:
affected_sets = [[ct, self.sets[self._fsset_key(ct)].membership(sample, perturb)]
for ct in np.arange(len(self.partitioner.sets))
if self.sets[self._fsset_key(ct)].membership(sample, perturb) > 0.0]
if len(affected_sets) == 0:
if sample < self.partitioner.lower_set().get_lower(perturb):
affected_sets.append([0, 1])
elif sample > self.partitioner.upper_set().get_upper(perturb):
affected_sets.append([len(self.sets) - 1, 1])
return affected_sets
[docs] def forecast(self, ndata, **kwargs):
time_displacement = kwargs.get("time_displacement",0)
window_size = kwargs.get("window_size", 1)
no_update = kwargs.get("no_update", False)
l = len(ndata)
ret = []
for k in np.arange(0, l):
sample = ndata[k]
if self.method == 'unconditional':
perturb = common.window_index(k + time_displacement, window_size)
elif self.method == 'conditional':
if not no_update:
perturb = self.conditional_perturbation_factors(sample)
else:
perturb = [[0, 1] for k in np.arange(len(self.partitioner.sets))]
affected_sets = self._affected_sets(sample, perturb)
numerator = []
denominator = []
if len(affected_sets) == 1:
ix = affected_sets[0][0]
aset = self.partitioner.ordered_sets[ix]
if aset in self.flrgs:
numerator.append(self.flrgs[aset].get_midpoint(self.sets, perturb[ix]))
else:
fuzzy_set = self.sets[aset]
numerator.append(fuzzy_set.get_midpoint(perturb[ix]))
denominator.append(1)
else:
for aset in affected_sets:
ix = aset[0]
fs = self.partitioner.ordered_sets[ix]
tdisp = perturb[ix]
if fs in self.flrgs:
numerator.append(self.flrgs[fs].get_midpoint(self.sets, tdisp) * aset[1])
else:
fuzzy_set = self.sets[fs]
numerator.append(fuzzy_set.get_midpoint(tdisp) * aset[1])
denominator.append(aset[1])
if sum(denominator) > 0:
pto = sum(numerator) / sum(denominator)
else:
pto = sum(numerator)
ret.append(pto)
if self.method == 'conditional' and not no_update:
self.forecasts.append(pto)
self.residuals.append(self.inputs[-1] - self.forecasts[-1])
self.inputs.append(sample)
self.inputs.pop(0)
self.forecasts.pop(0)
self.residuals.pop(0)
return ret
[docs] def forecast_interval(self, ndata, **kwargs):
time_displacement = kwargs.get("time_displacement", 0)
window_size = kwargs.get("window_size", 1)
l = len(ndata)
ret = []
for k in np.arange(0, l):
# print("input: " + str(ndata[k]))
tdisp = common.window_index(k + time_displacement, window_size)
affected_sets = [[self.sets[key], self.sets[key].membership(ndata[k], tdisp)]
for key in self.partitioner.ordered_sets
if self.sets[key].membership(ndata[k], tdisp) > 0.0]
if len(affected_sets) == 0:
affected_sets.append([common.check_bounds(ndata[k], self.partitioner, tdisp), 1.0])
upper = []
lower = []
if len(affected_sets) == 1:
aset = affected_sets[0][0]
if aset.name in self.flrgs:
lower.append(self.flrgs[aset.name].get_lower(tdisp))
upper.append(self.flrgs[aset.name].get_upper(tdisp))
else:
lower.append(aset.get_lower(tdisp))
upper.append(aset.get_upper(tdisp))
else:
for aset in affected_sets:
if aset[0].name in self.flrgs:
lower.append(self.flrgs[aset[0].name].get_lower(tdisp) * aset[1])
upper.append(self.flrgs[aset[0].name].get_upper(tdisp) * aset[1])
else:
lower.append(aset[0].get_lower(tdisp) * aset[1])
upper.append(aset[0].get_upper(tdisp) * aset[1])
ret.append([sum(lower), sum(upper)])
return ret
