卷积神经网络(CNN)、长短期记忆网络(LSTM)以及门控单元网络(GRU)是最常见的一类算法，在kaggle比赛中经常被用来做预测和回归。今天，我们就抛砖引玉，做一个简单的教程，如何用这些网络预测时间序列。因为是做一个简单教程，所以本例子中网络的层数和每层的神经元个数没有调试到最佳。根据不同的数据集，同学们可以自己调网络结构和具体参数。
1.环境搭建
我们运行的环境是下载anaconda，然后在里面安装keras，打开spyder运行程序即可。其中下载anaconda和安装keras的教程在我们另一个博客“用CNN做电能质量扰动分类(2019-03-28)”中写过了，这里就不赘述了。
2.数据集下载
下载时间序列数据集和程序。其中，网盘连接是：
https://pan.baidu.com/s/1TASK3gMZoDFvoE89LzR-5A，密码是“o0sl”。
“nihe.csv”是我自己做的一个时间序列的数据集，一共有1000行4列其中，1-3列可以认为是X，第4列认为是Y。我们现在要做的就是训练3个X和Y之间的关系，然后给定X去预测Y。
3.预测
把下载的nihe.csv文件放到spyder 的默认路径下，我的默认路径是“D:\Matlab2018a\42”，新建一个.py文件，把程序放进去，运行即可。
4.CNN,LSTM,GRU预测时间序列的程序
1）GRU的程序
#1. load dataset
from pandas import read_csv
dataset = read_csv('nihe.csv')
values = dataset.values
#2.tranform data to [0,1]
from sklearn.preprocessing importMinMaxScaler
scaler=MinMaxScaler(feature_range=(0, 1))
XY= scaler.fit_transform(values)
X= XY[:,0:3]
Y = XY[:,3]
#3.split into train and test sets
n_train_hours = 950
trainX = X[:n_train_hours, :]
trainY =Y[:n_train_hours]
testX = X[n_train_hours:, :]
testY =Y[n_train_hours:]
train3DX =trainX.reshape((trainX.shape[0], 1, trainX.shape[1]))
test3DX = testX.reshape((testX.shape[0],1, testX.shape[1]))
#4. Define Network
from keras.models importSequential
from keras.layers import Dense
from keras.layers.recurrentimport GRU
model = Sequential()
model.add(GRU(units=5,input_shape=(train3DX.shape[1],train3DX.shape[2]),return_sequences=True))
model.add(GRU(units=3))
model.add(Dense(units=4,kernel_initializer='normal',activation='relu'))
model.add(Dense(units=1,kernel_initializer='normal',activation='sigmoid'))
#最后输出层1个神经元和输出的个数对应
# 5. compile the network
model.compile(loss='mae',optimizer='adam')
# 6. fit the network
history =model.fit(train3DX,trainY, epochs=100, batch_size=10,validation_data=(test3DX,testY), verbose=2, shuffle=False)
# 7. evaluate the network
from matplotlib import pyplot
pyplot.plot(history.history['loss'],label='train')
pyplot.plot(history.history['val_loss'],label='test')
pyplot.legend()
pyplot.show()
#8. make a prediction and invertscaling for forecast
from pandas import concat
forecasttestY0 =model.predict(test3DX)
inv_yhat=np.concatenate((testX,forecasttestY0), axis=1)
inv_y =scaler.inverse_transform(inv_yhat)
forecasttestY = inv_y[:,3]
# calculate RMSE
from math import sqrt
from sklearn.metrics importmean_squared_error
actualtestY=values[n_train_hours:,3]
rmse = sqrt(mean_squared_error(forecasttestY,actualtestY))
print('Test RMSE: %.3f' % rmse)
#plot the testY and actualtestY
pyplot.plot(actualtestY,label='train')
pyplot.plot(forecasttestY,label='test')
pyplot.legend()
pyplot.show()
2）LSTM的程序
#1. load dataset
from pandas import read_csv
dataset = read_csv('nihe.csv')
values = dataset.values
#2.tranform data to [0,1]
from sklearn.preprocessing importMinMaxScaler
scaler=MinMaxScaler(feature_range=(0, 1))
XY= scaler.fit_transform(values)
X= XY[:,0:3]
Y = XY[:,3]
#3.split into train and test sets
n_train_hours = 950
trainX = X[:n_train_hours, :]
trainY =Y[:n_train_hours]
testX = X[n_train_hours:, :]
testY =Y[n_train_hours:]
train3DX =trainX.reshape((trainX.shape[0], 1, trainX.shape[1]))
test3DX = testX.reshape((testX.shape[0],1, testX.shape[1]))
#4. Define Network
from keras.models importSequential
from keras.layers import Dense
from keras.layers.recurrentimport LSTM
model = Sequential()
model.add(LSTM(units=5,input_shape=(train3DX.shape[1],train3DX.shape[2]),return_sequences=True))
model.add(LSTM(units=3))
model.add(Dense(units=4,kernel_initializer='normal',activation='relu'))
model.add(Dense(units=1,kernel_initializer='normal',activation='sigmoid'))
#最后输出层1个神经元和输出的个数对应
# 5. compile the network
model.compile(loss='mae',optimizer='adam')
# 6. fit the network
history =model.fit(train3DX,trainY, epochs=100, batch_size=10,validation_data=(test3DX,testY), verbose=2, shuffle=False)
# 7. evaluate the network
from matplotlib import pyplot
pyplot.plot(history.history['loss'],label='train')
pyplot.plot(history.history['val_loss'],label='test')
pyplot.legend()
pyplot.show()
#8. make a prediction and invertscaling for forecast
from pandas import concat
import numpy as np
forecasttestY0 =model.predict(test3DX)
#forecasttestY= np.expand_dims(a,axis=1)
inv_yhat=np.concatenate((testX,forecasttestY0), axis=1)
inv_y =scaler.inverse_transform(inv_yhat)
forecasttestY = inv_y[:,3]
# calculate RMSE
from math import sqrt
from sklearn.metrics importmean_squared_error
actualtestY=values[n_train_hours:,3]
rmse =sqrt(mean_squared_error(forecasttestY, actualtestY))
print('Test RMSE: %.3f' % rmse)
#plot the testY and actualtestY
pyplot.plot(actualtestY,label='train')
pyplot.plot(forecasttestY,label='test')
pyplot.legend()
pyplot.show()
3）CNN和LSTM的合并
#1. load dataset
from pandas import read_csv
dataset = read_csv('nihe.csv')
values = dataset.values
#2.tranform data to [0,1] 3个属性，第4个是待预测量
from sklearn.preprocessing importMinMaxScaler
scaler=MinMaxScaler(feature_range=(0, 1))
XY= scaler.fit_transform(values)
X= XY[:,0:3]
Y = XY[:,3]
#3.split into train and test sets
950个训练集，剩下的都是验证集
n_train_hours = 950
trainX = X[:n_train_hours, :]
trainY =Y[:n_train_hours]
testX = X[n_train_hours:, :]
testY =Y[n_train_hours:]
#LSTM的输入格式要3维，因此先做变换
train3DX =trainX.reshape((trainX.shape[0], 1, trainX.shape[1]))
test3DX =testX.reshape((testX.shape[0], 1, testX.shape[1]))
#4. Define Network
from keras.models importSequential
from keras.layers import Dense
from keras.layers.recurrentimport LSTM
from keras.layers.convolutionalimport Conv1D
from keras.layers.convolutionalimport MaxPooling1D
from keras.layers import Flatten
model = Sequential()
model.add(Conv1D(filters=10,kernel_size=1, padding='same', strides=1, activation='relu',input_shape=(1,3)))
model.add(MaxPooling1D(pool_size=1))
model.add(LSTM(units=3,return_sequences=True))
model.add(Flatten())
#可以把LSTM和Flatten删除,仅保留LSTM
#model.add(LSTM(units=3))
model.add(Dense(5,activation='relu'))
#在lstm层之后可以添加隐含层，也可以不加，直接加输出层
#model.add(Dense(units=4,kernel_initializer='normal',activation='relu'))
model.add(Dense(units=1,kernel_initializer='normal',activation='sigmoid'))
#最后输出层1个神经元和输出的个数对应
# 5. compile the network
model.compile(loss='mae',optimizer='adam')
# 6. fit the network
history =model.fit(train3DX,trainY, epochs=100, batch_size=10,validation_data=(test3DX,testY), verbose=2, shuffle=False)
# 7. evaluate the network
from matplotlib import pyplot
pyplot.plot(history.history['loss'],label='train')
pyplot.plot(history.history['val_loss'],label='test')
pyplot.legend()
pyplot.show()
#8. make a prediction and invertscaling for forecast
from pandas import concat
import numpy as np
forecasttestY0 =model.predict(test3DX)
inv_yhat=np.concatenate((testX,forecasttestY0), axis=1)
inv_y =scaler.inverse_transform(inv_yhat)
forecasttestY = inv_y[:,3]
# calculate RMSE
from math import sqrt
from sklearn.metrics importmean_squared_error
actualtestY=values[n_train_hours:,3]
rmse =sqrt(mean_squared_error(forecasttestY, actualtestY))
print('Test RMSE: %.3f' % rmse)
#plot the testY and actualtestY
pyplot.plot(actualtestY,label='train')
pyplot.plot(forecasttestY,label='test')
pyplot.legend()
pyplot.show()
