在 Keras 中使用 LSTM 进行多元时间序列预测（关于未来数据）

Question

So I have been using Keras to predict a multivariate time series. The dataset is a pollution dataset. The first column is what I want to predict and the remaining 7 are features. Dataset can be found here: https://github.com/sagarmk/Forecasting-on-Air-pollution-with-RNN-LSTM/blob/master/pollution.csv

So what I want to do is to perform the following code on a test set without the "pollution" column. Let's say that there is new data for the features but not the pollution. So

1. How do I train the model without test data? (model.fit())
2. How do I predict new pollution data without future data on pollution? (model.predict())

To make it simple the dataset could be initially split into a training and testing dataset in the beginning, where the "pollution" column is removed from he testing dataset?

See below a simple code. Here I simply import and process the dataset.

import numpy as np
import pandas as pd
import matplotlib as plt
import seaborn as sns
import plotly.express as px

# Import data
dataset = pd.read_csv("pollution.csv")
dataset = dataset.drop(['date'], axis = 1)
label, layer = pd.factorize(dataset['wnd_dir'])
dataset['wnd_dir'] = pd.DataFrame(label)
dataset=dataset.fillna(dataset.mean())
dataset.head()

After that I normalize the dataset.

from sklearn.preprocessing import MinMaxScaler

values = dataset.values
scaler = MinMaxScaler()
scaled = scaler.fit_transform(values)
scaled[0]

Then convert the normalized data into supervised form.

def to_supervised(dataset,dropNa = True,lag = 1):
    df = pd.DataFrame(dataset)
    column = []
    column.append(df)
    for i in range(1,lag+1):
        column.append(df.shift(-i))
    df = pd.concat(column,axis=1)
    df.dropna(inplace = True)
    features = dataset.shape[1]
    df = df.values
    supervised_data = df[:,:features*lag]
    supervised_data = np.column_stack( [supervised_data, df[:,features*lag]])
    return supervised_data

timeSteps = 2

supervised = to_supervised(scaled,lag=timeSteps)
pd.DataFrame(supervised).head()

Now the dataset is split and transformed so that the LSTM network can handle it.

features = dataset.shape[1]
train_hours = round(dataset.shape[0]*0.7)
X = supervised[:,:features*timeSteps]
y = supervised[:,features*timeSteps]

x_train = X[:train_hours,:]
x_test = X[train_hours:,:]
y_train = y[:train_hours]
y_test = y[train_hours:]

print(x_train.shape,x_test.shape,y_train.shape,y_test.shape)
#convert data to fit for lstm
#dimensions = (sample, timeSteps here it is 1, features )

x_train = x_train.reshape(x_train.shape[0], timeSteps, features)
x_test = x_test.reshape(x_test.shape[0], timeSteps, features)

print(x_train.shape,x_test.shape)

Here the model is trained

#define the model
from keras.models import Sequential
from keras.layers import Dense,LSTM

model = Sequential()
model.add( LSTM( 50, input_shape = ( timeSteps,x_train.shape[2]) ) )
model.add( Dense(1) )

model.compile( loss = "mae", optimizer = "adam")

history =  model.fit( x_train,y_train, validation_data = (x_test,y_test), epochs = 50 , batch_size = 72, verbose = 0, shuffle = False)
plt.pyplot.plot(history.history['loss'], label='train')
plt.pyplot.plot(history.history['val_loss'], label='test')
plt.pyplot.legend()
#plt.pyplot.yticks([])
#plt.pyplot.xticks([])
plt.pyplot.title("loss during training")
plt.pyplot.show()

Lastly I plot the training data along with the test data.

y_pred = model.predict(x_test)
x_test = x_test.reshape(x_test.shape[0],x_test.shape[2]*x_test.shape[1])

inv_new = np.concatenate( (y_pred, x_test[:,-7:] ) , axis =1)
inv_new = scaler.inverse_transform(inv_new)
final_pred = inv_new[:,0]

plt.pyplot.figure(figsize=(20,10))
plt.pyplot.plot(dataset['pollution'])
plt.pyplot.plot([None for i in dataset['pollution']] + [x for x in final_pred])
plt.pyplot.show()

Answer 1

Predicting results with your neural network should be as simple as the below line of code. Passing new data that is in the same format as training data.

prediction = model.predict(features)

Do you have any code that you can provide? What issue are you running into?

Answer 2

Ie when the "test" dataset only consists of 8 feature columns and no column for the price?

It looks like you are asking a feature engeering question. You real dataset have nan value in different column which make predict failed, right?

In this case, you can take commom solution:

fill nan value by the median/mean of correspoding column in trainset.

For example, you can fill future price by the median/mean of recently 14 days(aggregation length) prices of each product.

UPDATE

When making future prediction, there may be a lot of features only have history(without plan). In traditional machine learning, if you want to predict a target depend on all feature, you need predict those future of features first.

But by LSTM, you can make prediction all in one, check time_series#multi-output_models