Projet_scientific

import numpy as np import scipy as scp import matplotlib.pyplot as plt from sklearn import neighbors from sklearn.model_selection import TimeSeriesSplit from sklearn.metrics import mean_squared_error import spacepy.time as spt import spacepy.omni as om from sklearn import tree from sklearn import svm from sklearn.ensemble import RandomForestRegressor from sklearn.datasets import make_regression import graphviz

On charge une bonne période de données

ticks = spt.tickrange('1980-01-01T00:00:00', '2017-01-01T00:00:00', deltadays = 1./24.) d = om.get_omni(ticks)

Kp = d['Kp'] Bz = d['BzIMF'] V = d['velo'] N = d['dens'] dates = d['ticks'].UTC

for d in (Kp, Bz, V, N): plt.figure() plt.plot(dates, d)

input_data = np.stack((Bz,V,N)).T output_data = Kp[:,None]

from datetime import datetime test_min, test_max = datetime(2003,10,22), datetime(2003,11,5) in_test = np.logical_and(dates > test_min, dates < test_max) train_index, = np.nonzero(np.logical_not(in_test)) test_index, = np.nonzero(in_test)

X_train, X_test = input_data[train_index], input_data[test_index] y_train, y_test = output_data[train_index], output_data[test_index]

plt.plot(dates[test_index], y_test) plt.figure() plt.plot(dates[train_index], y_train)

knn = neighbors.KNeighborsRegressor(10, weights='distance') model = knn.fit(X_train, y_train)

p_test = model.predict(X_test) plt.plot(dates[test_index], y_test) plt.plot(dates[test_index], p_test) plt.legend(("Kp réel", "Prédiction")) print("RMSE =" , np.sqrt(mean_squared_error(y_test, p_test)))

p_test = model.predict(X_test) plt.plot(dates[test_index], p_test) plt.plot(dates[test_index], y_test) plt.legend(("Kp réel", "Prédiction")) print(np.shape(p_test)) np.sqrt(mean_squared_error(y_test, p_test))

tss = TimeSeriesSplit(n_splits=100)

Évaluation de l'erreur du model:

errors = [] m_dates = [] for train_index, test_index in tss.split(input_data, output_data): X_train, X_test = input_data[train_index], input_data[test_index] y_train, y_test = output_data[train_index], output_data[test_index] dates_test = dates[test_index] model = knn.fit(X_train, y_train) p_test = model.predict(X_test) error = np.sqrt(mean_squared_error(y_test, p_test)) errors.append(error) m_dates.append(dates_test[len(test_index)//2]) #plt.figure() #plt.plot(dates_test, y_test) #plt.plot(dates_test, p_test)

plt.plot(m_dates,errors)

Évolution de l'erreur moyenne en fonction du temps.

En 1994/1995, deux nouveaux satellites (ACE,WIND) sont mis en services.

Ce qui explique la forte baisse de l'erreur (moins de trous dans les données).

On change la période de données après la partie 1 parce qu'on sait la carence de ACE et WIND dans l'estimateur

ticks_change = spt.tickrange('1980-01-01T00:00:00', '2017-01-01T00:00:00', deltadays = 1./24.) d1 = om.get_omni(ticks_change)

Kp1 = d1['Kp'] Bz1 = d1['BzIMF'] V1 = d1['velo'] N1 = d1['dens'] dates1 = d1['ticks'].UTC

for d in (Kp1, Bz1, V1, N1): plt.figure() plt.plot(dates1, d)

input_data1 = np.stack((Bz1,V1,N1)).T output_data1 = Kp1[:,None]

Sélection d'une période de test:

test_min, test_max = datetime(2005,10,22), datetime(2005,11,5) in_test = np.logical_and(dates1 > test_min, dates1 < test_max) train_index, = np.nonzero(np.logical_not(in_test)) test_index, = np.nonzero(in_test)

X_train1, X_test1 = input_data1[train_index], input_data1[test_index] y_train1, y_test1 = output_data1[train_index], output_data1[test_index]

plt.plot(dates1[test_index], y_test1) print(np.shape(dates1[test_index])) plt.figure() plt.plot(dates1[train_index], y_train1)

model_svm = svm.LinearSVR().fit(X_train1, y_train1.ravel())

s_test = model_svm.predict(X_test1) plt.plot(dates1[test_index], y_test1) plt.plot(dates1[test_index], s_test) plt.legend(("Kp réel", "Prédiction selon svm")) print("RMSE:", np.sqrt(mean_squared_error(y_test1, s_test)))

tss = TimeSeriesSplit(n_splits=100)

Évaluation de l'erreur du model:

errors = [] m_dates = [] for train_index, test_index in tss.split(input_data1, output_data1): X_train1, X_test1 = input_data1[train_index], input_data1[test_index] y_train1, y_test1 = output_data1[train_index], output_data1[test_index] dates_test = dates1[test_index] model_svm = svm.LinearSVR().fit(X_train1, y_train1.ravel()) s_test = model_svm.predict(X_test1) error = np.sqrt(mean_squared_error(y_test1, s_test)) errors.append(error) m_dates.append(dates_test[len(test_index)//2]) #plt.figure() #plt.plot(dates_test, y_test) #plt.plot(dates_test, p_test)

plt.plot(m_dates,errors)

On charge une bonne période de données

ticks = spt.tickrange('2012-01-01T00:00:00', '2017-01-01T00:00:00', deltadays = 1./24.) d = om.get_omni(ticks) Kp = d['Kp'] Bz = d['BzIMF'] V = d['velo'] N = d['dens'] dates = d['ticks'].UTC

input_data = np.stack((Bz,V,N)).T output_data = Kp[:,None]

Sélection d'une période de test:

from datetime import datetime test_min, test_max = datetime(2015,10,22), datetime(2015,10,23) in_test = np.logical_and(dates > test_min, dates < test_max) train_index, = np.nonzero(np.logical_not(in_test)) test_index, = np.nonzero(in_test)

X_train, X_test = input_data[train_index], input_data[test_index] y_train, y_test = output_data[train_index], output_data[test_index]

plt.plot(dates[test_index], y_test) print(np.shape(y_test)) plt.figure() plt.plot(dates[train_index], y_train)

SVR = svm.SVR(kernel='rbf', tol=1e-5,cache_size=7000,gamma='scale') model2 = SVR.fit(X_train, y_train.ravel())

pred = model2.predict(X_test) plt.plot(dates[test_index], y_test) plt.plot(dates[test_index], pred) plt.legend(("Kp réel", "Prédiction")) print(np.shape(pred)) print(np.shape(y_test)) print("RMSE=",np.sqrt(mean_squared_error(y_test, pred)))

plt.plot(dates[test_index], y_test) plt.plot(dates[test_index], pred) plt.legend(("Kp réel", "Prédiction")) print(np.shape(pred)) print(np.shape(y_test)) print("RMSE=",np.sqrt(mean_squared_error(y_test, pred)))

tss = TimeSeriesSplit(n_splits=1000)

Évaluation de l'erreur du model:

plt.plot(m_dates,errors)

Évolution de l'erreur moyenne en fonction du temps.

En 1994/1995, deux nouveaux satellites (ACE,WIND) sont mis en services.

Ce qui explique la forte baisse de l'erreur (moins de trous dans les données).

On charge une bonne période de données

ticks = spt.tickrange('1980-01-01T00:00:00', '2017-01-01T00:00:00', deltadays = 1./24.) d = om.get_omni(ticks)

Kp = d['Kp'] Bz = d['BzIMF'] V = d['velo'] N = d['dens'] dates = d['ticks'].UTC

input_data = np.stack((Bz,V,N)).T output_data = Kp[:,None]

Sélection d'une période de test:

from datetime import datetime test_min, test_max = datetime(2014,10,22), datetime(2014,11,5) in_test = np.logical_and(dates > test_min, dates < test_max) train_index, = np.nonzero(np.logical_not(in_test)) test_index, = np.nonzero(in_test)

X_train, X_test = input_data[train_index], input_data[test_index] y_train, y_test = output_data[train_index], output_data[test_index]

plt.plot(dates[test_index], y_test) print(np.shape(dates[test_index]))

clf = tree.DecisionTreeRegressor() clf = clf.fit(X_test,y_test )

dot_data=tree.export_graphviz(clf) graph = graphviz.Source(dot_data)

graph

regr = tree.DecisionTreeRegressor() model=regr.fit(X_train,y_train) p_test = model.predict(X_test) plt.plot(dates[test_index], y_test) plt.plot(dates[test_index], p_test) plt.legend(("Kp réel", "Prédiction")) print("RMSE=", np.sqrt(mean_squared_error(y_test, p_test)))

tss = TimeSeriesSplit(n_splits=100)

Évaluation de l'erreur du model:

errors = [] m_dates = [] for train_index, test_index in tss.split(input_data, output_data): X_train, X_test = input_data[train_index], input_data[test_index] y_train, y_test = output_data[train_index], output_data[test_index] dates_test = dates[test_index] model = regr.fit(X_train, y_train) p_test = model.predict(X_test) error = np.sqrt(mean_squared_error(y_test, p_test)) errors.append(error) m_dates.append(dates_test[len(test_index)//2]) #plt.figure() #plt.plot(dates_test, y_test) #plt.plot(dates_test, p_test)

plt.plot(m_dates,errors)

On charge une bonne période de données

ticks = spt.tickrange('1980-01-01T00:00:00', '2017-01-01T00:00:00', deltadays = 1./24.) d = om.get_omni(ticks)

Kp = d['Kp'] Bz = d['BzIMF'] V = d['velo'] N = d['dens'] dates = d['ticks'].UTC

input_data = np.stack((Bz,V,N)).T output_data = Kp[:,None]

Sélection d'une période de test:

X_train, X_test = input_data[train_index], input_data[test_index] y_train, y_test = output_data[train_index], output_data[test_index]

plt.plot(dates[test_index], y_test) print(np.shape(dates[test_index]))

regr = RandomForestRegressor(random_state=1, n_estimators=10) model=regr.fit(X_train, y_train.ravel()) p_test = model.predict(X_test) plt.plot(dates[test_index], y_test) plt.plot(dates[test_index], p_test) plt.legend(("Kp réel", "Prédiction"))

print("RMSE= ",np.sqrt(mean_squared_error(y_test, p_test)))

tss = TimeSeriesSplit(n_splits=100)

Évaluation de l'erreur du model:

errors = [] m_dates = [] for train_index, test_index in tss.split(input_data, output_data): X_train, X_test = input_data[train_index], input_data[test_index] y_train, y_test = output_data[train_index], output_data[test_index] dates_test = dates[test_index] model = regr.fit(X_train, y_train.ravel()) p_test = model.predict(X_test) error = np.sqrt(mean_squared_error(y_test, p_test)) errors.append(error) m_dates.append(dates_test[len(test_index)//2]) #plt.figure() #plt.plot(dates_test, y_test) #plt.plot(dates_test, p_test)

plt.plot(m_dates,errors)

vnale / projet_scientific Goto Github PK