def optimize(postfix):
# sg.utils.redirect(sys.stdout, "gridopt_output_%s.txt" % postfix)
user_id = 55864860
(dataset, test) = load_prediction.prepare_datasets(user_id)
day = 24
freerun = day
today = 4600
# [len_data, res_size, leak, input, bias, spectral,
# seed, ridge, tmp_sm, load_sm]
train_hours = 336
datas = \
[sg.utils.Normalizer(dataset[today-train_hours:today+day-freerun,:], axis=0)
for today in (1000, 2000, 3000, 4000)]
input_data = []
for data in datas:
temps, loads = zip(*data.normalized)
input_data.append([np.array((temps[24:], loads[:-24], loads[24:])).T])
reservoir = Oger.nodes.LeakyReservoirNode(output_dim=400,
leak_rate=1,
input_scaling=0.5,
bias_scaling=0.75,
spectral_radius=1,
reset_states=False)
readout = Oger.nodes.RidgeRegressionNode(ridge_param = 0.001)
flow = Oger.nodes.FreerunFlow(reservoir + readout,
freerun_steps = freerun,
external_input_range= \
np.array([0, 1]))
# gridsearch_parameters = {reservoir: {'_instance': range(5),
# 'spectral_radius': [0.6, 0.8, 1],
# 'input_scaling': [0.1, 0.5, 0.9],
# 'bias_scaling': [0.1, 0.5, 0.9],
# 'leak_rate': [0.1, 0.5, 0.9]},
# readout: {'_instance': range(5),
# 'ridge_param': [0.1, 0.5, 0.9]}}
gridsearch_parameters = {reservoir: {'_instance': range(20)},
readout: {'ridge_param': [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]}}
print "gridsearch_parameters = " + str(gridsearch_parameters)
optimizer = Oger.evaluation.Optimizer(gridsearch_parameters,
Oger.utils.nrmse)
optimizer.grid_search([[], input_data], flow,
cross_validate_function=Oger.evaluation.leave_one_out)
return (optimizer, reservoir)
import numpy as np import Oger, mdp import matplotlib.pyplot as plt import scikits.timeseries as ts import pywt import sg.utils from sg.data.sintef.create_full_temp_data import data as read_temperatures import sg.data.sintef.userloads as ul import load_prediction from static import StaticNode user_id = 55864860 (dataset, test) = load_prediction.prepare_datasets(user_id, False) #day = 24 #today = random.randint(1000, dataset.shape[0]-day*2) #today = 4600 #See if we can predict 24 times based on instances, learned from the training set. data_raw = sg.utils.Normalizer(dataset, axis=0) data = data_raw.normalized[:2**14,1] # One year is 365*24 = 8760 datapoints. If we round down to 8192, we will get # the maximum amount of scales for the decomposition (13), i.e. math.pow(2,13) # The number of levels/scales determine how far we look back. level = 4
def optimize(postfix):
# sg.utils.redirect(sys.stdout, "gridopt_output_%s.txt" % postfix)
user_id = 55864860
(dataset, test) = load_prediction.prepare_datasets(user_id)
day = 24
freerun = day
today = 4600
# [len_data, res_size, leak, input, bias, spectral,
# seed, ridge, tmp_sm, load_sm]
train_hours = 336
datas = \
[sg.utils.Normalizer(dataset[today-train_hours:today+day-freerun,:], axis=0)
for today in (1000, 2000, 3000, 4000)]
input_data = []
for data in datas:
temps, loads = zip(*data.normalized)
input_data.append([np.array((temps[24:], loads[:-24], loads[24:])).T])
reservoir = Oger.nodes.LeakyReservoirNode(output_dim=400,
leak_rate=1,
input_scaling=0.5,
bias_scaling=0.75,
spectral_radius=1,
reset_states=False)
readout = Oger.nodes.RidgeRegressionNode(ridge_param=0.001)
flow = Oger.nodes.FreerunFlow(reservoir + readout,
freerun_steps = freerun,
external_input_range= \
np.array([0, 1]))
# gridsearch_parameters = {reservoir: {'_instance': range(5),
# 'spectral_radius': [0.6, 0.8, 1],
# 'input_scaling': [0.1, 0.5, 0.9],
# 'bias_scaling': [0.1, 0.5, 0.9],
# 'leak_rate': [0.1, 0.5, 0.9]},
# readout: {'_instance': range(5),
# 'ridge_param': [0.1, 0.5, 0.9]}}
gridsearch_parameters = {
reservoir: {
'_instance': range(20)
},
readout: {
'ridge_param': [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
}
}
print "gridsearch_parameters = " + str(gridsearch_parameters)
optimizer = Oger.evaluation.Optimizer(gridsearch_parameters,
Oger.utils.nrmse)
optimizer.grid_search(
[[], input_data],
flow,
cross_validate_function=Oger.evaluation.leave_one_out)
return (optimizer, reservoir)