from sklearn.svm import SVR from utility import Utility from datetime import datetime import sys #Get the commamd line arguments directoryName = "Datasets/" profileName = sys.argv[1] datasetFileName = directoryName + profileName + "_time_interaction_rate.csv" daysOfHorizon = 30 daysOfDepth = 14 horizon = 24 * daysOfHorizon #7 days ahead depth = 24 * daysOfDepth #30 days util = Utility.SeriesUtility() # Step 1 - Convert the dataset into pandas series series = util.convertDatasetsToSeries(datasetFileName) # Step 2 - Resample the series (to hourly) resampledSeries = util.resampleSeriesMean(series, "H") del series # Todo - Train based on the recent data only. yearsOfData = 5 recentCount = yearsOfData * 365 * 24 + horizon filteredSeries = util.filterRecent(resampledSeries, recentCount) del resampledSeries # Step 3 - Scale the series normalizedSeries = util.scaleSeries(filteredSeries)
]
profileNames = ["Audi", "Benz", "Dodge", "Ferrari", "Jeep"]
# For plotting
seriesList = []
seriesNames = []
# Parameters for training
daysOfHorizon = 4
daysOfDepth = 30
horizon = 24 * daysOfHorizon
depth = 24 * daysOfDepth
yearsOfData = 5
# Util Object
globalUtil = Utility.SeriesUtility()
# Preprocess and sample the series( Series Intersection)
resampledSeries = []
for i in range(len(datasetFileNames)):
# Convert the dataset into pandas series
series = globalUtil.convertDatasetsToSeries(datasetFileNames[i])
# Resample the series (to hourly)
resampled = globalUtil.resampleSeries(series, "H")
resampledSeries.append(resampled)
del series
# Find intersection
processedSeries = globalUtil.intersect(resampledSeries)
def predictSeries(seriesName):
# Dataset
directoryName = "Datasets/"
datasetFileName = directoryName + seriesName + "_time_interaction.csv"
util = Utility.SeriesUtility()
# Step 1 - Convert the dataset into pandas series
series = util.convertDatasetsToSeries(datasetFileName)
# Step 2 - Re-sample the series (to hourly)
resampledSeries = util.resampleSeriesSum(series, "H")
del series
# Step 3 - Filter recent data
yearsOfData = 3
recentCount = yearsOfData * 365 * 24 + horizon
filteredSeries = util.filterRecent(resampledSeries, recentCount)
del resampledSeries
# Step 4 - Scale the series
normalizedSeries = util.scaleSeries(filteredSeries)
del filteredSeries
# Step 5 - Split into training and testing
trainingSeries, testingSeries = util.splitIntoTrainingAndTestingSeries(
normalizedSeries, horizon)
# Step 6 - Form the feature and target vectors for training
featureIntervalList = []
# # 24 hour interval
period = 60
for i in range(period, 0, -1):
t = -24 * i
featureIntervalList.append(pd.Timedelta(hours=t)) # T
targetIntervalList = [pd.Timedelta(hours=0)]
featureTrainingVectors, targetTrainingVectors = tsi.TimeSeriesIntervalProcessor(
trainingSeries, featureIntervalList,
targetIntervalList).getProcessedData()
featureTrainingVectors = np.hstack((np.ones(
(featureTrainingVectors.shape[0], 1)), featureTrainingVectors))
# Step 7 - Train the network
networkSize = int(featureTrainingVectors.shape[0] / 10)
util.trainESNWithoutTuning(size=networkSize,
featureVectors=featureTrainingVectors,
targetVectors=targetTrainingVectors,
initialTransient=50,
inputConnectivity=1.0,
reservoirConnectivity=0.3,
inputScaling=0.5,
reservoirScaling=0.5,
spectralRadius=0.79,
leakingRate=0.30,
learningMethod=Utility.LearningMethod.Batch)
# Step 8 - Predict the future
predictedSeries = util.predictFutureWithFeatureInterval(
trainingSeries, featureIntervalList, horizon)
# Step 9 - De-scale the series
actualSeries = util.descaleSeries(testingSeries)
predictedSeries = util.descaleSeries(predictedSeries)
return actualSeries, predictedSeries
def predictSeries(seriesName):
# Dataset
directoryName = "Datasets/"
profileName = seriesName
datasetFileName = directoryName + profileName + "_time_interaction.csv"
util = Utility.SeriesUtility()
# Step 1 - Convert the dataset into pandas series
series = util.convertDatasetsToSeries(datasetFileName)
# Step 2 - Re-sample the series (to hourly)
resampledSeries = util.resampleSeriesSum(series, "H")
del series
# Step 3 - Filter recent data
recentCount = yearsOfData * 365 * 24 + horizon
filteredSeries = util.filterRecent(resampledSeries, recentCount)
del resampledSeries
# Step 4 - Scale the series
normalizedSeries = util.scaleSeries(filteredSeries)
del filteredSeries
# Step 5 - Split into training, validation and testing series
trainingSeries, testingSeries = util.splitIntoTrainingAndTestingSeries(
normalizedSeries, horizon)
availableSeries = trainingSeries
trainingSeries, validationSeries = util.splitIntoTrainingAndValidationSeries(
trainingSeries, trainingSetRatio=0.9)
networkParameters['size'] = int(
trainingSeries.shape[0] /
10) # One-tenth of length of the training set
#networkParameters['size'] = 256
# Step 6 - Form the feature and target vectors for training
bestDepth, bestFeaturesIndices, bestFeatures, targetVectors = util.getBestFeatures(
trainingSeries,
validationSeries,
networkParameters,
Utility.FeatureSelectionMethod.Pattern_Analysis,
args={'threshold': 0.5})
# Step 7 - Tune the leaking rate by brute force
optimalLeakingRate = util.getBestLeakingRate(bestFeaturesIndices,
bestDepth, bestFeatures,
targetVectors, trainingSeries,
validationSeries,
networkParameters)
# Step 7 - Tune and Train the network
util.trainESNWithoutTuning(
size=networkParameters['size'],
featureVectors=bestFeatures,
targetVectors=targetVectors,
initialTransient=networkParameters['initialTransient'],
inputConnectivity=networkParameters['inputConnectivity'],
reservoirConnectivity=networkParameters['reservoirConnectivity'],
inputScaling=networkParameters['inputScaling'],
reservoirScaling=networkParameters['reservoirScaling'],
spectralRadius=networkParameters['spectralRadius'],
leakingRate=optimalLeakingRate)
# Step 8 - Predict the future
predictedSeries = util.predict(util.esn, availableSeries, bestDepth,
horizon, bestFeaturesIndices)
# Step 9 - De-scale the series
actualSeries = util.descaleSeries(testingSeries)
predictedSeries = util.descaleSeries(predictedSeries)
return actualSeries, predictedSeries
