Keras 没有做出好的预测

Question

两个月前，我开始使用 keras 以获得泵模式，以便在其他软件中使用它。

我不知道为什么我获得的图案与真实图案无关。我尝试通过在数据集中建立一些特征（输入）以及更多输入，但它无法正常工作。 结果看起来像这样：

地点：

• 蓝色：数据集（我试图“近似”的真实数据）
• 橙色：预测

数据集是一个时间序列

here是带有数据集的csv文件

代码如下：

import numpy
import matplotlib.pyplot as plt
import pandas
import math

from keras.models import Sequential
from keras.layers import Dense, LSTM, Dropout
from sklearn.preprocessing import MinMaxScaler
from sklearn.metrics import mean_squared_error
from keras.regularizers import l2, activity_l2


def create_dataset(dataset, look_back=1):
    dataX, dataY = [], []
    for i in range(len(dataset) - look_back - 1):
        a = dataset[i:(i + look_back), 0:4] 
        dataX.append(a)
        dataY.append(dataset[i + look_back, 4]) 
    return numpy.array(dataX), numpy.array(dataY)



# fix random seed for reproducibility
seed=7
numpy.random.seed(seed)


# load dataset
dataframe = pandas.read_csv('datos_horarios.csv', engine='python') 
dataset = dataframe.values


# normalizar el dataset
scaler = MinMaxScaler(feature_range=(0, 1))
dataset = scaler.fit_transform(dataset)

#split data into train data and test data
train_size = int(len(dataset) * 0.67) 
test_size = len(dataset) - train_size
train, test = dataset[0:train_size, :], dataset[train_size:len(dataset), :]

# reshape to X=t y Y=t+1
look_back = 1
trainX, trainY = create_dataset(train, look_back)  
testX, testY = create_dataset(test, look_back)

# reshape inputs to be [samples, time steps, features]
trainX = numpy.reshape(trainX, (trainX.shape[0], look_back, 4))
testX = numpy.reshape(testX, (testX.shape[0], look_back, 4))
# create and adjust LSTM network

model = Sequential()
model.add(Dropout(0.3, input_shape=(look_back,4))) 
model.add(LSTM(6, input_shape=(look_back,4), W_regularizer=l2(0.001))) 
model.add(Dense(10)) 
model.add(Dense(1))
model.compile(loss='mean_squared_error', optimizer='adam' ,momentum=0.99)
history= model.fit(trainX, trainY,validation_split=0.33, nb_epoch=250, batch_size=32)


# Plot
plt.plot(history.history['loss'])
plt.plot(history.history['val_loss'])
plt.title('model loss')
plt.ylabel('loss')
plt.xlabel('epochs')
plt.legend(['training', 'validation'], loc='upper right')
plt.show()

# make predictions
trainPredict = model.predict(trainX)
testPredict = model.predict(testX)
print(trainPredict)

numero_inputs=4
inp=numero_inputs-1
# Get something which has as many features as dataset
trainPredict_extended = numpy.zeros((len(trainPredict),numero_inputs+1))
# Put the predictions there
trainPredict_extended[:,inp+1] = trainPredict[:,0]
# Inverse transform it and select the 3rd column.
trainPredict = scaler.inverse_transform(trainPredict_extended)[:,inp+1]

# Get something which has as many features as dataset
testPredict_extended = numpy.zeros((len(testPredict),numero_inputs+1))
# Put the predictions there
testPredict_extended[:,inp+1] = testPredict[:,0]
# Inverse transform it and select the 3rd column.
testPredict = scaler.inverse_transform(testPredict_extended)[:,inp+1]

trainY_extended = numpy.zeros((len(trainY),numero_inputs+1))
trainY_extended[:,inp+1]=trainY
trainY=scaler.inverse_transform(trainY_extended)[:,inp+1]

testY_extended = numpy.zeros((len(testY),numero_inputs+1))
testY_extended[:,inp+1]=testY
testY=scaler.inverse_transform(testY_extended)[:,inp+1]
# Calcular error medio cuadratico
trainScore = math.sqrt(mean_squared_error(trainY, trainPredict))
print('Train Score: %.2f RMSE' % (trainScore))
testScore = math.sqrt(mean_squared_error(testY, testPredict))
print('Test Score: %.2f RMSE' % (testScore))
# add train predictions to the plot

trainPredictPlot = numpy.empty_like(dataset)
trainPredictPlot[:, :] = numpy.nan
trainPredictPlot[look_back:len(trainPredict)+look_back, 0] = trainPredict

# add test predictions to the plot
testPredictPlot = numpy.empty_like(dataset)
testPredictPlot[:, :] = numpy.nan
testPredictPlot[len(trainPredict)+(look_back*2)+1:len(dataset)-1, 0] = testPredict

# Plot real data and training and test predictions
serie,=plt.plot(scaler.inverse_transform(dataset)[:,numero_inputs])  #invierto muestras en formato (0,1) a valores reales y los ploteo
entrenamiento,=plt.plot(trainPredictPlot[:,0],linestyle='--')  #ploteo las predicciones de entrenamiento
prediccion_test,=plt.plot(testPredictPlot[:,0],linestyle='--')
plt.ylabel(' (m3)')
plt.xlabel('h')
plt.legend([serie,entrenamiento,prediccion_test],['Time series','Training','Prediction'], loc='upper right')
plt.show()

关于如何解决此问题的任何想法？或者，至少，问题是什么？

按列输入：

1. 一天中的时间（每半小时），转换为十进制。
2. 星期几（1-星期一...7-星期日）
3. 一年中的月份 (1-12)
4. 一个月中的某天 (1-31)

输出：

5. 抽水（立方米）

编辑 使用 @a_guest 的代码，并更改一些参数，例如 epoch 数或 history 值，结果非常好：

Answer 1

不是答案，但我在这里分享我获得以下结果的代码：

请注意，网络参数是任意选择的，即未优化。也就是说，您很可能通过改变这些参数来获得更好的结果。此外，改变history（或您的情况下为look_back）的值可能会对预测质量产生重大影响。

from keras.models import Sequential
from keras.layers import Dense
import matplotlib.pyplot as plt
import numpy

numpy.random.seed(12)

history = 96


def generate_data():
    data = numpy.loadtxt('datos_horarios.csv', delimiter=',', dtype=float)
    # Normalize data.
    data[:, -1] /= numpy.max(data[:, -1])
    train_test_data = []
    for i in xrange(data.shape[0] - history - 1):
        # Include the reference value here, will be extracted later.
        train_test_data.append(data[i:i+history+1, -1].flatten())
    return numpy.array(train_test_data)

train_test_data = generate_data()
# Shuffle data set in order to randomly select training and test data.
numpy.random.shuffle(train_test_data)

n_samples = train_test_data.shape[0]
n_train_samples = int(0.8 * n_samples)

train_data = train_test_data[:n_train_samples, :-1]
train_data_reference = train_test_data[:n_train_samples, -1][:, None]

test_data = train_test_data[n_train_samples:, :-1]
test_data_reference = train_test_data[n_train_samples:, -1]

print 'Tranining data: ', train_data
print 'Reference values: ', train_data_reference


model = Sequential()
model.add(Dense(history, input_dim=history, activation='sigmoid'))
model.add(Dense(history/2, activation='sigmoid'))
model.add(Dense(history/4, activation='sigmoid'))
model.add(Dense(history/8, activation='sigmoid'))
model.add(Dense(history/16, activation='sigmoid'))
model.add(Dense(1, activation='sigmoid'))
model.compile(optimizer='rmsprop', loss='mean_squared_error', metrics=['accuracy'])
model.summary()

model.fit(train_data, train_data_reference, shuffle=True, nb_epoch=200, batch_size=10)

# Use the complete data set to see the network performance.
# Regenerate data set because it was shuffled before.
train_test_data = generate_data()
test_data_predicted = model.predict(train_test_data[:, :-1]).flatten()
test_data_reference = train_test_data[:, -1]

relative_deviation = test_data_predicted/test_data_reference - 1.0
print 'Relative deviation: ', relative_deviation

plt.figure()
plt.plot(range(len(test_data_reference)), test_data_reference, 'b-', label='reference')
plt.plot(range(len(test_data_predicted)), test_data_predicted, 'r--', label='predicted')
plt.xlabel('test case #')
plt.ylabel('predictions')
plt.title('Reference values vs predicted values')
plt.legend()

plt.figure()
plt.plot(range(len(test_data_predicted)), relative_deviation, 'bx', label='relative deviation')
plt.xlabel('test case #')
plt.ylabel('relative deviation')
plt.title('Relative deviation of predicted values (predicted / reference - 1)')
plt.legend()

plt.show()