def clusterRegression(stations, startDate, endDate, \
features, clusterFeatures=None, \
nclusters=1, ranseed=666, \
targetVar='TempMax', \
lag=1, order=0, scale=False, verbose=False):
# build regression model to predict a variable for a single
# station using training data from multiple stations
# between startdate and enddate. Uses a "Taylor expansion"
# by combining information from several days (higher order
# time derivatives)
#
# stations: a list of station codes, the first entry is
# the station for which forecast is generated
# features: a list of variables to use as predictors
# *** if a feature string contains a ":" it is parsed as
# an interaction between two features ...
# *** features in interaction terms pre-scaled!
# clusterFeatures: subset of features with respect to which
# k-means clustering is applied before training
# regression models
# nclusters: number of clusters to compute
# lag: the number of days in the future to forecast
# order: the number of days in the past to include
# (also maximum order of time derivative)
import wUCluster as Clust
reload(Clust)
import numpy as np
import wUUtils as Util
from sklearn import preprocessing
from sklearn import linear_model
# load target variable data
target = Util.loadDailyVariableRange(stations[0], startDate, endDate, \
targetVar, castFloat=True)
# shift vector by lag
target = target[lag:]
# load feature data
featureData = []
prescalers = []
for station in stations:
for feature in features:
# check if feature contains an interaction
if ':' in feature:
feat1 = feature.split(':')[0]
feat2 = feature.split(':')[1]
fd1 = Util.loadDailyVariableRange(station, startDate, endDate, \
feat1, castFloat=True)
fd2 = Util.loadDailyVariableRange(station, startDate, endDate, \
feat2, castFloat=True)
prescaler1 = preprocessing.StandardScaler().fit(fd1)
fd1 = prescaler1.transform(fd1)
prescaler2 = preprocessing.StandardScaler().fit(fd2)
fd2 = prescaler2.transform(fd2)
# save prescaler objects (for prediction)
prescalers.append([prescaler1,prescaler2])
# compute interaction
fd = (np.array(fd1)*np.array(fd2)).tolist()
else:
fd = Util.loadDailyVariableRange(station, startDate, endDate, \
feature, castFloat=True)
prescalers.append(None)
# shorten vector by lag
fd = fd[:(-lag)]
featureData.append(fd)
# add in "derivative" terms
for ideriv in range(1,order+1):
ncols = len(stations)*len(features)
for ii in range(ncols):
# print("Adding " + str(ideriv) + " derivative of " + feature[jfeat])
fd = np.diff(featureData[ii],n=ideriv)
featureData.append(fd)
# shorten vectors to length of highest order derivative
nrows = len(featureData[-1])
for column in range(len(featureData)):
featureData[column] = featureData[column][-nrows:]
target = target[-nrows:]
# apply k-means clustering
if clusterFeatures is not None:
classes, clusterParams = Clust.clusterFeatureData(featureData, stations, features, \
clusterFeatures, nclusters, \
ranseed)
classes, featureClusters = Clust.assignClustersAllFeatures(featureData, clusterParams)
targetClusters = []
for cl in range(nclusters):
targetClusters.append([t for i,t in enumerate(target) if classes[i] == cl])
else:
# everything is one cluster
classes = range(len(target))
featureClusters = [featureData]
targetClusters = [target]
clusterParams = { 'nclusters': 1 }
# train separate regression model for each cluster
regrs = []
scalers = []
for icl in range(nclusters):
# convert features and target to arrays
featureClusters[icl] = (np.array(featureClusters[icl])).T
targetClusters[icl] = np.array(targetClusters[icl])
scaler = None
if scale:
scaler = preprocessing.StandardScaler().fit(featureClusters[icl])
featureClusters[icl] = scaler.transform(featureClusters[icl])
scalers.append(scaler)
regr = linear_model.LinearRegression()
regr.fit(featureClusters[icl], targetClusters[icl])
regrs.append(regr)
print('Cluster %d, nrows %d, R^2 %f' \
% (icl, \
len(targetClusters[icl]), \
regr.score(featureClusters[icl],targetClusters[icl])) )
if verbose:
print("Regression coefficients:")
print(" intercept" + ":\t" + str(regr.intercept_))
column = 0
for ideriv in range(order+1):
print(" " + str(ideriv) + "th derivative:")
for jj, station in enumerate(stations):
print(" Station: " + station)
for ii, feature in enumerate(features):
print(" " + feature + ":\t" + str(regr.coef_[column]))
column += 1
# save model parameters
modelParams = {
'stations': stations, \
'startDate': startDate, \
'endDate': endDate, \
'targetVar': targetVar, \
'features': features, \
'regrs': regrs, \
'clusterParams': clusterParams, \
'classes': classes, \
'lag': lag, \
'order': order, \
'scale': scale, \
'scalers': scalers, \
'prescalers': prescalers}
return featureData, target, modelParams
def clusterRegressionPredict(modelParams, startDate, endDate, actual=True):
# predict targetVar for a single station using
# previously generated regression model
import wUCluster as Clust
reload(Clust)
import numpy as np
import wUUtils as Util
# extract city and feature data
stations = modelParams['stations']
targetVar = modelParams['targetVar']
features = modelParams['features']
regrs = modelParams['regrs']
clusterParams = modelParams['clusterParams']
nclusters = clusterParams['nclusters']
lag = modelParams['lag']
order = modelParams['order']
scale = modelParams['scale']
prescalers = modelParams['prescalers']
scalers = modelParams['scalers']
# build list of dates in datetime format
date_list = Util.dateList(startDate, endDate)
date_list = date_list[(lag+order):]
# if actual data available
if actual:
# load target variable data
target = Util.loadDailyVariableRange(stations[0], startDate, endDate, \
targetVar, castFloat=True)
# "baseline" model is predicted target same as value on prediction day
baseline = target[order:(-lag)]
baseline = np.array(baseline)
# shift vector by lag
target = target[lag:]
target = np.array(target)
else:
target = None
# load feature data
featureData = []
idata = 0
for station in stations:
for feature in features:
# check if feature contains an interaction
if ':' in feature:
feat1 = feature.split(':')[0]
feat2 = feature.split(':')[1]
fd1 = Util.loadDailyVariableRange(station, startDate, endDate, \
feat1, castFloat=True)
fd2 = Util.loadDailyVariableRange(station, startDate, endDate, \
feat2, castFloat=True)
# rescale factors in interaction
prescaler1, prescaler2 = prescalers[idata]
fd1 = prescaler1.transform(fd1)
fd2 = prescaler2.transform(fd2)
# compute interaction
fd = (np.array(fd1)*np.array(fd2)).tolist()
else:
fd = Util.loadDailyVariableRange(station, startDate, endDate, \
feature, castFloat=True)
# shorten vector by lag
fd = fd[:(-lag)]
featureData.append(fd)
# increment feature counter
idata += 1
# add in "derivative" terms
for ideriv in range(1,order+1):
ncols = len(stations)*len(features)
for ii in range(ncols):
# print("Adding " + str(ideriv) + " derivative of " + feature[jfeat])
fd = np.diff(featureData[ii],n=ideriv)
featureData.append(fd)
# shorten vectors to length of highest order derivative
nrows = len(featureData[-1])
for column in range(len(featureData)):
featureData[column] = featureData[column][-nrows:]
if actual:
target = target[-nrows:]
# allocate features to clusters
if clusterParams['nclusters'] > 1:
classes, featureClusters = Clust.assignClustersAllFeatures(featureData, clusterParams)
dateClusters = []
for icl in range(nclusters):
dateClusters.append([t for i,t in enumerate(date_list) if classes[i] == icl])
if actual:
targetClusters = []
for icl in range(nclusters):
targetClusters.append([t for i,t in enumerate(target) if classes[i] == icl])
else:
# everything is one cluster
classes = range(len(target))
featureClusters = [featureData]
dateClusters = [date_list]
if actual:
targetClusters = [target]
preds = []
RMSE = []
R2 = []
for icl in range(nclusters):
# convert features and target to arrays
featureClusters[icl] = (np.array(featureClusters[icl])).T
if scale:
scaler = scalers[icl]
featureClusters[icl] = scaler.transform(featureClusters[icl])
regr = regrs[icl]
preds.append(regr.predict(featureClusters[icl]))
if actual:
targetClusters[icl] = np.array(targetClusters[icl])
print('Cluster %d, %d rows:' % (icl,len(dateClusters[icl])) )
r2 = regrs[icl].score(featureClusters[icl],targetClusters[icl])
print(' R^2_mean:' + '\t' + str(r2))
rmse = np.sqrt(((preds[icl] - targetClusters[icl])**2).mean())
print(' RMSE:\t' + '\t' + str(rmse))
RMSE.append(rmse)
R2.append(r2)
# assemble predictions into one list
date_list_mixed = np.concatenate(dateClusters).tolist()
pred_mixed = np.concatenate(preds).tolist()
pred = [pr for (d,pr) in sorted(zip(date_list_mixed,pred_mixed))]
if actual:
rmse = np.sqrt(((np.array(pred) - np.array(target))**2).mean())
print('\nOverall performance:')
print(' RMSE:' + '\t' + str(rmse))
modelPerf = {'RMSE': RMSE, 'R2': R2, 'RMSE_total': rmse }
else:
modelPerf = None
return date_list, pred, target, featureData, classes, modelPerf
