def detectDoubleExponentialSmoothingAnomalies(series, y, alpha, beta, threshold=2, mae = 0, deviation= 0, bound=IS_UPPER_BOUND, calculateScore=False, debug=False, returnAnomaliesOnly= True):
yhat = double_exponential_smoothing(y, alpha, beta)
# calculate confidence intervals for smoothed values
if( mae==0 and deviation ==0 ):
mae = mean_absolute_error(y, yhat)
if calculateScore==True or debug == True:
mape = mean_absolute_percentage_error(y, yhat)
logging.info("Mean absolute error is "+str(mae)+", Mean absolute percentage_error is "+str(mape)+"%")
if (mae ==0 and deviation == 0 ):
deviation = np.std(y - yhat)
lower_bond = yhat - (mae + threshold * deviation)
upper_bond = yhat + (mae + threshold * deviation)
if debug == True:
logging.info(lower_bond)
logging.info(upper_bond)
logging.info(y)
ts=[]
adata=[]
anomalies=[]
myshape = series.shape
nrow = myshape[0]
i=0
for i in range(nrow):
isAnomaly = False
if (not returnAnomaliesOnly):
ts.append(series.index[i])
adata.append(series.iloc[i,0])
if bound==IS_UPPER_BOUND:
if y[i] > upper_bond[i]:
if returnAnomaliesOnly:
ts.append(series.index[i])
adata.append(y[i])
isAnomaly = True
elif bound==IS_LOWER_BOUND:
if y[i] < lower_bond[i]:
if returnAnomaliesOnly:
ts.append(series.index[i])
adata.append(y[i])
isAnomaly = True
else:
if y[i] > upper_bond[i] or y[i] < lower_bond[i]:
if returnAnomaliesOnly:
ts.append(series.index[i])
adata.append(y[i])
isAnomaly = True
if returnAnomaliesOnly:
if isAnomaly:
anomalies.append(True)
else:
anomalies.append(isAnomaly)
#return mae, deviation, addHeader(ts,adata)
if calculateScore:
return mae, deviation, ts, adata, anomalies, mape
return mae, deviation, ts, adata, anomalies
def trainDoubleExponentialSmoothingParameters(y, alphas, betas):
minmap = -1
alpha = 0
beta = 0
for a in alphas:
for b in betas:
yhat = double_exponential_smoothing(y, a, b)
mape = mean_absolute_percentage_error(y, yhat)
if (minmap == -1):
minmap = mape
alpha = a
beta = b
elif minmap > mape:
minmap = mape
alpha = a
beta = b
return alpha, beta
def calculateDoubleExponentialSmoothingParameters( series, alpha=0.95, beta=0.05,threshold = 2, calculateScore=False):
y= series.y
yhat = double_exponential_smoothing(y.values, alpha, beta)
# calculate confidence intervals for smoothed values
mae = mean_absolute_error(y.values, yhat)
deviation = np.std(y.values - yhat)
mape = 0
if calculateScore:
mape = mean_absolute_percentage_error(y.values, yhat)
logging.info("Mean absolute error is "+str(mae)+", Mean absolute percentage_error is "+str(mape)+"%")
hlower_bound = yhat - (mae + threshold * deviation)
hupper_bound = yhat + (mae + threshold * deviation)
if calculateScore:
return hlower_bound[len(hlower_bound)-1], hupper_bound[len(hupper_bound)-1], mape
return hlower_bound[len(hlower_bound)-1], hupper_bound[len(hupper_bound)-1]
