def _build_model(
self, observed_time_series, trend='local', seasonal_periods=12, ar_order=1
):
if trend == 'semilocal':
trend = sts.SemiLocalLinearTrend(
observed_time_series=observed_time_series, name='trend'
)
else:
trend = sts.LocalLinearTrend(
observed_time_series=observed_time_series, name='trend'
)
mods_list = [trend]
if str(seasonal_periods).isdigit():
seasonal = tfp.sts.SmoothSeasonal(
period=seasonal_periods,
observed_time_series=observed_time_series,
name='seasonal',
frequency_multipliers=[1, 2, 3],
)
mods_list.append(seasonal)
if str(ar_order).isdigit():
autoregressive = sts.Autoregressive(
order=int(ar_order),
observed_time_series=observed_time_series,
name='autoregressive',
)
mods_list.append(autoregressive)
model = sts.Sum(mods_list, observed_time_series=observed_time_series)
return model
def build_model(observed_time_series):
trend = sts.LocalLinearTrend(observed_time_series=observed_time_series)
seasonal = tfp.sts.Seasonal(num_seasons=12,
observed_time_series=observed_time_series)
model = sts.Sum([trend, seasonal],
observed_time_series=observed_time_series)
return model
def build_model(Risk_Factors):
# Seasonality effect if relevant: daily data assumed here, datetime index assumed to be comprised of business days
day_of_week_effect = sts.Seasonal(
num_seasons=5,
observed_time_series=np.array(
Risk_Factors[Risk_Factors.columns[1]]).astype(np.float32),
name='day_of_week_effect')
# effect of each explicative factor
data = np.array(Risk_Factors[Risk_Factors.columns[2:]]).astype(np.float32)
other_effects = sts.LinearRegression(design_matrix=tf.reshape(
data - np.mean(data), (-1, data.shape[1])),
name='other_effects')
# auto-regressive effect
autoregressive = sts.Autoregressive(
order=1,
observed_time_series=np.array(
Risk_Factors[Risk_Factors.columns[1]]).astype(np.float32),
name='autoregressive')
model = sts.Sum([day_of_week_effect, other_effects, autoregressive],
observed_time_series=np.array(
Risk_Factors[Risk_Factors.columns[1]]).astype(
np.float32))
return model
def build_model(observed_time_series):
# trend = sts.LocalLinearTrend(observed_time_series=observed_time_series)
seasonal_daily = tfp.sts.Seasonal(
num_seasons=48,
num_steps_per_season=1,
observed_time_series=observed_time_series,
name='day_of_week_effect')
seasonal_weekly = tfp.sts.Seasonal(
num_seasons=5,
num_steps_per_season=48 * 5,
observed_time_series=observed_time_series,
name='week_of_year_effect')
model = sts.Sum([seasonal_daily, seasonal_weekly],
observed_time_series=observed_time_series)
return model
def build_model(observed_time_series):
day_of_week_effect = sts.Seasonal(
num_seasons=7,
observed_time_series=observed_time_series,
name='day_of_week_effect')
month_of_year_effect = sts.Seasonal(
num_seasons=12,
num_steps_per_season=num_days_per_month,
observed_time_series=observed_time_series,
name='month_of_year_effect')
autoregressive = sts.Autoregressive(
order=1,
observed_time_series=observed_time_series,
name='autoregressive')
model = sts.Sum([day_of_week_effect, month_of_year_effect, autoregressive],
observed_time_series=observed_time_series)
return model
def build_model(self, day=True, month=True, res=True, ext=False):
if day:
self.day_of_week = self.day_of_week_effect()
if month:
self.month_of_yr = self.month_of_yr_effect()
if ext:
self.external = self.external_effect()
if res:
self.residue = self.residue_effect()
# get rid of None in the list
list_of_effects = [
self.day_of_week, self.month_of_yr, self.external, self.residue
]
list_of_effects = list(filter(None.__ne__, list_of_effects))
self.model = sts.Sum(components=list_of_effects,
observed_time_series=self.obs)
def cal_loss(training_data):
#设置全局默认图形
tf.reset_default_graph()
#遵循加法模型，设置趋势
trend = sts.LocalLinearTrend(observed_time_series=observed_time_series)
#设置季节性
seasonal = tfp.sts.Seasonal(
num_seasons=12, observed_time_series=observed_time_series)
#模型拟合,之所以用sum，而不是我们在建模中常见的fit定义，是因为，
#模型时间序列为加法模型，有如上文提到的趋势，季节性，周期性等成分相加
#默认的先验分布为正态（normal）
ts_model = sts.Sum([trend, seasonal], observed_time_series=observed_time_series)
#构建变分损失函数和后验
with tf.variable_scope('sts_elbo', reuse=tf.AUTO_REUSE):
elbo_loss, variational_posteriors = tfp.sts.build_factored_variational_loss(
ts_model,observed_time_series=training_data)
return ts_model,elbo_loss,variational_posteriors
def build_model(observed_time_series):
trend = sts.LocalLinearTrend(observed_time_series=observed_time_series)
seasonal = tfp.sts.Seasonal(
num_seasons=int(len(dataset)/janela), observed_time_series=observed_time_series)
model = sts.Sum([trend, seasonal], observed_time_series=observed_time_series)
return model
