-
Notifications
You must be signed in to change notification settings - Fork 0
/
svr_model_selection.py
144 lines (106 loc) · 5.12 KB
/
svr_model_selection.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
# -*- coding: utf-8 -*-
"""
Created on Wed May 25 10:17:49 2016
@author: Magnus Dahl
"""
import pandas as pd
import datetime as dt
import numpy as np
from sklearn import linear_model
from sklearn.svm import SVR
from sklearn.preprocessing import StandardScaler
from sklearn.cross_validation import cross_val_predict
from model_selection import gen_all_combinations, rmse, mae, mape
import sql_tools as sq
import ensemble_tools as ens
#%% SVR experinment
ts = ens.gen_hourly_timesteps(dt.datetime(2016,1,26,1), dt.datetime(2016,4,1,0))
X = pd.read_pickle('48h60h168h_lagged_X.pkl') # run model_selection_ext_horizon to generate these files
y = pd.read_pickle('prod_to_gowith.pkl')
# add more predictor data:
for v in ['Tout', 'vWind', 'hum', 'sunRad']:
X[v] = ens.load_ens_mean_avail_at10_series(v, ts[0], ts[-1], pointcode=71699)
#X['weekdays'] = [t.weekday() for t in ts]
def h_hoursbefore(timestamp, h):
return timestamp + dt.timedelta(hours=-h)
most_recent_avail_prod = sq.fetch_production(h_hoursbefore(ts[0], 24),\
h_hoursbefore(ts[-1], 24))
for i, t, p48 in zip(range(len(most_recent_avail_prod)), ts, X['prod48hbefore']):
if t.hour > 8 or t.hour == 0:
most_recent_avail_prod[i] = p48
X['prod24or48hbefore'] = most_recent_avail_prod
##
X_scaler = StandardScaler(copy=True, with_mean=True, with_std=True).fit(X)
X_scaled = X_scaler.transform(X)
y_scaler = StandardScaler(copy=True, with_mean=True, with_std=True).fit(y)
y_scaled = y_scaler.transform(y)
#%%
svr_rmses = {}
svr_poly_rmses = {}
#Cs = np.linspace(14,18,10)
#gams = np.linspace(0.003, 0.004, 10)
Cs = np.linspace(35,55, 10)
gams = np.linspace(0.002, 0.005, 10)
# SVR with RBF kernel
for C in Cs:
for gamma in gams:
svr_rbf = SVR(kernel='rbf', C=C, gamma=gamma)
y_rbf = cross_val_predict(svr_rbf, X_scaled, y_scaled, cv=10)
svr_rmses[(C,gamma)] = rmse(y_scaler.inverse_transform(y_rbf)-y)
#%%
svr_rbf = SVR(kernel='rbf', C=35, gamma=0.00266)
y_rbf = cross_val_predict(svr_rbf, X_scaled, y_scaled, cv=10)
svr_rmse_lone = rmse(y_scaler.inverse_transform(y_rbf)-y)
svr_mae_lone = mae(y_scaler.inverse_transform(y_rbf)-y)
svr_mape_lone = mape(y_scaler.inverse_transform(y_rbf)-y,y)
lr = linear_model.LinearRegression()
y_lin = cross_val_predict(lr, X_scaled, y_scaled, cv=10)
#%% load ensemble data
def gen_ens_dfs(ts_start, ts_end, varnames, timeshifts, pointcode=71699):
""" timeshifts must be integer number of hours. Posetive values only,
dataframe contains columns with the variables minus their value
'timeshift' hours before. """
df = pd.DataFrame()
df_s = [pd.DataFrame() for i in range(25)]
for timeshift in timeshifts:
prod_before = sq.fetch_production(h_hoursbefore(ts_start, timeshift),\
h_hoursbefore(ts_end, timeshift))
for df in df_s:
df['prod%ihbefore'%timeshift] = prod_before
for v in varnames:
ens_data = ens.load_ens_avail_at10_series(ts_start, ts_end, v, pointcode=71699)
ens_data_before = ens.load_ens_avail_at10_series(h_hoursbefore(ts_start, timeshift),\
h_hoursbefore(ts_end, timeshift), v, pointcode=71699)
diff = ens_data - ens_data_before
for i in range(ens_data.shape[1]):
df_s[i]['%s%ihdiff%i'%(v,timeshift, i)] = diff[:,i]
for v in varnames:
ens_data = ens.load_ens_avail_at10_series(ts_start, ts_end, v, pointcode=71699)
for i in range(ens_data.shape[1]):
df_s[i]['%s%i'%(v, i)] = ens_data[:,i]
for df in df_s:
df['prod24or48hbefore'] = most_recent_avail_prod
return df_s
ens_X_data = gen_ens_dfs(ts[0], ts[-1], ['Tout', 'vWind', 'hum', 'sunRad'],[48, 60, 168])
svr_rbf.fit(X_scaled,y_scaled)
ens_y_data = [svr_rbf.predict(X_scaler.transform(x)) for x in ens_X_data]
ens_ydata_scaled = [y_scaler.inverse_transform(yy) for yy in ens_y_data]
#%% Test on blind period!
ts_test = ens.gen_hourly_timesteps(dt.datetime(2016,4,1,1), dt.datetime(2016,5,1,0))
X_test = pd.read_pickle('48h60h168h_lagged_X_test.pkl')
X_test = X_test.ix[:, X_test.columns !='prod'] # run model_selection_ext_horizon to generate these files
y_test = pd.read_pickle('prod_to_gowith_test.pkl')
# add more predictor data:
for v in ['Tout', 'vWind', 'hum', 'sunRad']:
X_test[v] = ens.load_ens_mean_avail_at10_series(v, ts_test[0], ts_test[-1], pointcode=71699)
most_recent_avail_prod_test = sq.fetch_production(h_hoursbefore(ts_test[0], 24),\
h_hoursbefore(ts_test[-1], 24))
for i, t, p48 in zip(range(len(most_recent_avail_prod_test)), ts_test, X_test['prod48hbefore']):
if t.hour > 8 or t.hour == 0:
most_recent_avail_prod[i] = p48
X_test['prod24or48hbefore'] = most_recent_avail_prod_test
#%%
test_pred = svr_rbf.predict(X_scaler.transform(X_test))
test_err = y_scaler.inverse_transform(test_pred) - y_test
print "SVR model performance"
print rmse(test_err), mae(test_err), mape(test_err,y_test)