# -*- coding: utf-8 -*- """XGBoostWB_Forecasting_Using_Hybrid_DL_Framework_Pm2.5_(1,6) """ import sys sys.version #Import Libraries import itertools import pandas as pd import numpy as np import os #import matplotlib #import matplotlib.pyplot as plt import random # %matplotlib inline import shutil from random import shuffle from sklearn.multioutput import MultiOutputRegressor import xgboost as xgb #TF Version tf.__version__ num_periods_output = 6 #to predict num_periods_input=1 #input ALL_Test_Data=[] ALL_Test_Prediction=[] """## preprocessing""" from sklearn.preprocessing import MinMaxScaler from sklearn.preprocessing import Normalizer def preprocessing(df_,num_features): cols=df_.columns #print('shape before',df_.shape) #print('date',df_['Date'].dtype) df_['Date'] = pd.to_datetime(df_['Date'], format='%Y-%m-%d') #print('date',df_['Date'].dtype) df_['DayofWeek']=df_['Date'].dt.dayofweek df_['Week']=df_['Date'].dt.week df_['dayofyear']=df_['Date'].dt.dayofyear Train=df_[(df_.year==2010) | (df_.year==2011)| (df_.year==2012)| (df_.year==2013)] #Train=df_[(df_.year==2010) | (df_.year==2011)| (df_.year==2012)] Test=df_[(df_.year==2014)] Train=Train.fillna(Train.mean()) Test=Test.fillna(Test.mean()) ################################################encoding######################## Train=Train[['pm2.5','No','year', 'month', 'day', 'hour', 'DEWP', 'TEMP', 'PRES', 'cbwd', 'Iws', 'Is', 'Ir']] cbwd=Train.pop('cbwd') Train.loc[:,'cbwd_cv']=(cbwd=='cv')*1.0 Train.loc[:,'cbwd_NE']=(cbwd=='NE')*1.0 Train.loc[:,'cbwd_NW']=(cbwd=='NW')*1.0 Train.loc[:,'cbwd_SE']=(cbwd=='SE')*1.0 Train=Train.values Train = Train.astype('float32') ################################################################################# Test=Test[['pm2.5','No','year', 'month', 'day', 'hour', 'DEWP', 'TEMP', 'PRES', 'cbwd', 'Iws', 'Is', 'Ir']] cbwd=Test.pop('cbwd') Test.loc[:,'cbwd_cv']=(cbwd=='cv')*1.0 Test.loc[:,'cbwd_NE']=(cbwd=='NE')*1.0 Test.loc[:,'cbwd_NW']=(cbwd=='NW')*1.0 Test.loc[:,'cbwd_SE']=(cbwd=='SE')*1.0 Test=Test.values Test = Test.astype('float32') ################################################################################# Number_Of_Features=num_features #split=num_periods_output+num_periods_input ############################# Normalization on train ############# PM_Train=Train[:,0] Train=np.delete(Train,[0],1) #normalizing data print('Len of training ',Train) Train = Train.astype('float32') normalizer = MinMaxScaler().fit(Train) Train=normalizer.transform(Train) PM_Train=np.reshape(PM_Train,(len(PM_Train),1)) Train=np.append(PM_Train,Train, axis=1) ############################################ TRAIN windows ################################## end=len(Train) start=0 next=0 x_batches=[] y_batches=[] count=0 #print('len',len(Train)) limit=num_periods_output+num_periods_input while start+(limit)<=end: next=start+num_periods_input x_batches.append(Train[start:next,:]) y_batches.append(Train[next:next+num_periods_output,0]) start=start+1 x_batches=np.asarray(x_batches) #print('x-------------',len(x_batches)) x_batches = x_batches.reshape(-1, num_periods_input, Number_Of_Features) y_batches=np.asarray(y_batches) #print('y=======',len(y_batches)) y_batches = y_batches.reshape(-1, num_periods_output, 1) #print('len x_batches ',len(x_batches)) ###########################################TEST Normalization################################## PM_Test=Test[:,0] Test=np.delete(Test,[0],1) Test = Test.astype('float32') Test=normalizer.transform(Test) PM_Test=np.reshape(PM_Test,(len(PM_Test),1)) Test=np.append(PM_Test,Test, axis=1) #------------------ ############################################ TEST windows ################################## end_test=len(Test) start_test=0 next_test=0 x_testbatches=[] y_testbatches=[] #print('len',len(Train)) while start_test+(limit)<=end_test: next_test=start_test+num_periods_input x_testbatches.append(Test[start_test:next_test,:]) y_testbatches.append(Test[next_test:next_test+num_periods_output,0]) start_test=start_test+num_periods_output x_testbatches=np.asarray(x_testbatches) #print('x----------',len(x_testbatches)) x_testbatches = x_testbatches.reshape(-1, num_periods_input, Number_Of_Features) y_testbatches=np.asarray(y_testbatches) #print('y====',len(y_testbatches)) y_testbatches = y_testbatches.reshape(-1, num_periods_output, 1) print('len Test',len(Test)) print('len xTestbatches',len(x_testbatches)) return x_batches, y_batches, x_testbatches, y_testbatches data_All=pd.DataFrame() x_batches_Full=[] y_batches_Full=[] X_Test_Full=[] Y_Test_Full=[] range_list = [1] data=pd.read_csv('/GBRT-for-TSF/Data/Multivariate/PM2_5.csv') header=list(data.columns.values) #print(header) data=pd.DataFrame(data,columns=header) x_batches_Full, y_batches_Full,X_Test_Full,Y_Test_Full=preprocessing(data,16) #---------------------shuffle minibatches X and Y together------------------------------------- #print(len(x_batches_Full),' length of all file : ',len(y_batches_Full)) combined = list(zip(x_batches_Full, y_batches_Full)) random.shuffle(combined) shuffled_batch_features, shuffled_batch_y = zip(*combined) #xgboost part #print(len(x_batches_Full)) All_Training_Instances=[] #=============== change each window into Instance ================================= for i in range(0,len(shuffled_batch_features)): hold=[] temp=[] for j in range(0,len(shuffled_batch_features[i])): #print(len(hold)) if j==(len(shuffled_batch_features[i])-1): hold=np.concatenate((hold, shuffled_batch_features[i][j][:]), axis=None) else: hold=np.concatenate((hold, shuffled_batch_features[i][j][0]), axis=None) #print(len(hold)) All_Training_Instances.append(hold) #print(len(All_Training_Instances[0])) #=============== change each window into Instance ================================= All_Testing_Instances=[] #print(len(X_Test_Full)) for i in range(0,len(X_Test_Full)): hold=[] temp=[] for j in range(0,len(X_Test_Full[i])): #print(len(hold)) if j==(len(X_Test_Full[i])-1): hold=np.concatenate((hold, X_Test_Full[i][j][:]), axis=None) else: hold=np.concatenate((hold, X_Test_Full[i][j][0]), axis=None) All_Testing_Instances.append(hold) #prediction=multioutput.predict(All_Testing_Instances) #print(len(All_Testing_Instances[0])) #===========================calling MultiOutput XGoost========================= All_Testing_Instances=np.reshape(All_Testing_Instances, (len(All_Testing_Instances),len(All_Testing_Instances[0]))) Y_Test_Full=np.reshape(Y_Test_Full, (len(Y_Test_Full),num_periods_output)) #========== reshape train ============================== All_Training_Instances=np.reshape(All_Training_Instances, (len(All_Training_Instances),len(All_Training_Instances[0]))) shuffled_batch_y=np.reshape(shuffled_batch_y, (len(shuffled_batch_y),num_periods_output)) #print(All_Training_Instances.shape) model=xgb.XGBRegressor(learning_rate =0.02, n_estimators=420, max_depth=3, min_child_weight=1, gamma=0.0, subsample=0.95, colsample_bytree=0.95, scale_pos_weight=0.9, seed=42,silent=False) multioutput=MultiOutputRegressor(model).fit(All_Training_Instances,shuffled_batch_y) print('Fitting Done!') prediction=multioutput.predict(All_Testing_Instances) print('prediction ',prediction.shape) print('test ',Y_Test_Full.shape) MSE=np.mean(( prediction- Y_Test_Full)**2) print('RMSE: ',MSE**0.5) MAE=np.mean(np.abs( prediction- Y_Test_Full)) print('MAE: ',MAE)
一、数据预处理
cols=df_.columns #取出列数
df_['Date'] = pd.to_datetime(df_['Date'], format='%Y-%m-%d') #将数据类型转换成datetime格式
df_['DayofWeek']=df_['Date'].dt.dayofweek #DAYOFWEEK函数返回日期的工作日索引值，即星期日为1，星期一为2，星期六为7。
Train=Train.fillna(Train.mean()) #用平均值填充缺失值
Train=df_[(df_.year==2010) | (df_.year==2011)| (df_.year==2012)| (df_.year==2013)]
Test=df_[(df_.year==2014)] #2014年作为测试数据，其他作为训练数据
cbwd=Train.pop('cbwd') #返回pop删除的值并赋值给对象，原列表改变。pop()默认为最后一个元素，即pop(-1)，pop(index)指定索引
Train.loc[:,'cbwd_cv']=(cbwd=='cv')*1.0 #loc定位，[x:y,a:b] 前者为行范围，后者为列范围 后面没懂？？
Test=Test.values #values取出除了行列名称的所有值（此处用处？）
Test = Test.astype('float32') #astype 更改数据类型
PM_Train=Train[:,0] #取出所有行和第0列的数值
Train=np.delete(Train,[0],1) #numpy.delete(arr, obj, axis) 处理Train矩阵，在[0]处处理，axis=1代表按列删除。delete用法
normalizer = MinMaxScaler().fit(Train) #归一化处理 MinMaxScaler为其中一种归一化类型
Train=normalizer.transform(Train) #transform函数（不是很懂 后面再看看）
x_batches=np.asarray(x_batches) #asarray函数：将数据类型转化为数组型
x_batches = x_batches.reshape(-1, num_periods_input, Number_Of_Features)
#reshape函数：改变行列，-1表示系统自动计算行/列，reshape(m,n)代表转换成m行n列
x_batches_Full, y_batches_Full,X_Test_Full,Y_Test_Full=preprocessing(data,16) #对数据进行分块与预处理
combined = list(zip(x_batches_Full, y_batches_Full)) #用zip函数将x，y两个数组中的每一个元素逐一组成新的元组
random.shuffle(combined) #shuffle 打乱
shuffled_batch_features, shuffled_batch_y = zip(*combined) #zip(*)函数 zip(*)是一种转置，zip(*)函数利用 * 号操作符，可以将元组解压为列表
二、 xgboost
hold=np.concatenate((hold, shuffled_batch_features[i][j][:]), axis=None) # concatenate函数 高效对大规模数据进行拼接
model=xgb.XGBRegressor(learning_rate =0.02, #进行xgboost，传入参数
n_estimators=420,
max_depth=3,
min_child_weight=1,
gamma=0.0,
subsample=0.95,
colsample_bytree=0.95,
scale_pos_weight=0.9,
seed=42,silent=False)
multioutput=MultiOutputRegressor(model).fit(All_Training_Instances,shuffled_batch_y)
prediction=multioutput.predict(All_Testing_Instances)
MSE=np.mean(( prediction- Y_Test_Full)**2) # MSE：均方误差，作为回归问题的损失函数
print('RMSE: ',MSE**0.5)
MAE=np.mean(np.abs( prediction- Y_Test_Full)) # 均绝对误差：预测值f(x)和真实值y之间距离的均值
print('MAE: ',MAE)
