用于温度时间序列预测的 LSTM 神经网络

Question

I'm learning to work with neural.networks applied to time-series so I tuned and LSTM example that I found to make predictions of daily temperature data. However, I found that the results are extremely poor as is shown in the image. (I only predict the last 92 days in order to save time for now).

This is the code I implemented. The data are 3 column dataframe (minimum, maximum and mean daily temperatures), but I only employ one of the columns at one time.

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from statsmodels.tools.eval_measures import rmse
from sklearn.preprocessing import MinMaxScaler
from keras.preprocessing.sequence import TimeseriesGenerator
from keras.models import Sequential
from keras.layers import Dense
from keras.layers import LSTM
from keras.layers import Dropout
import warnings
warnings.filterwarnings("ignore")

input_file2 = "TemperaturasCampillos.txt"
seriesT = pd.read_csv(input_file2,sep = "\t", decimal = ".", names = ["Minimas","Maximas","Medias"])
seriesT[seriesT==-999]=np.nan

date1 = '2010-01-01'
date2 = '2010-09-01'
date3 = '2020-05-17'
date4 = '2020-12-31'
mydates = pd.date_range(date2, date3).tolist()
seriesT['Fecha'] = mydates
seriesT.set_index('Fecha',inplace=True)  # Para que los índices sean fechas y así se ponen en el eje x de forma predeterminada
seriesT.index = seriesT.index.to_pydatetime()

df =  seriesT.drop(seriesT.columns[[1, 2]], axis=1)  # df.columns is zero-based pd.Index
n_input = 92
train, test = df[:-n_input], df[-n_input:]

scaler = MinMaxScaler()
scaler.fit(train)
train = scaler.transform(train)
test = scaler.transform(test)


#n_input = 365
n_features = 1
generator = TimeseriesGenerator(train, train, length=n_input, batch_size=1)
model = Sequential()
model.add(LSTM(200, activation='relu', input_shape=(n_input, n_features)))
model.add(Dropout(0.15))
model.add(Dense(1))
model.compile(optimizer='adam', loss='mse')
model.fit_generator(generator,epochs=150)

#create an empty list for each of our 12 predictions
#create the batch that our model will predict off of
#save the prediction to our list
#add the prediction to the end of the batch to be used in the next prediction

pred_list = []

batch = train[-n_input:].reshape((1, n_input, n_features))

for i in range(n_input):   
    pred_list.append(model.predict(batch)[0]) 
    batch = np.append(batch[:,1:,:],[[pred_list[i]]],axis=1)

df_predict = pd.DataFrame(scaler.inverse_transform(pred_list),                           
                          index=df[-n_input:].index, columns=['Prediction'])
df_test = pd.concat([df,df_predict], axis=1)

plt.figure(figsize=(20, 5))
plt.plot(df_test.index, df_test['Minimas'])
plt.plot(df_test.index, df_test['Prediction'], color='r')
plt.legend(loc='best', fontsize='xx-large')
plt.xticks(fontsize=18)
plt.yticks(fontsize=16)
plt.show()

As you can see if you click in the image link, I get a predict too smoothed, good to see the seasonality but is not what I am looking forward. In addition, I tried to add more layers to the neural.network shown, so the.network looks something like:

#n_input = 365
n_features = 1
generator = TimeseriesGenerator(train, train, length=n_input, batch_size=1)
model = Sequential()
model.add(LSTM(200, activation='relu', input_shape=(n_input, n_features)))
model.add(LSTM(128, activation='relu'))
model.add(LSTM(256, activation='relu'))
model.add(LSTM(128, activation='relu'))
model.add(LSTM(64, activation='relu'))
model.add(LSTM(n_features, activation='relu'))
model.add(Dropout(0.15))
model.add(Dense(1))
model.compile(optimizer='adam', loss='mse')
model.fit_generator(generator,epochs=100)

but I get this error:

ValueError : Input 0 is incompatible with layer lstm_86: expected ndim=3, found ndim=2

Of course, as the model has a bad performance I cannot assure that out-of-sample predictions would be accurate. Why I cannot implement more layers to the.network? How could I improve the performance?

Answer 1

You are missing one argument: return_sequences.

When you have more than one LSTM layer, you should set it to TRUE. Because, otherwise, that layer will only output the last hidden state. Add it to each LSTM layer.

model.add(LSTM(128, activation='relu', return_sequences=True))

About poor performance: my guess it is because of you have low amount of data for this application (the data seems pretty noise), add layers won't help too much.