LSTM NN给出了到达时间的裤子预测，我哪里做错了？

如何解决LSTM NN给出了到达时间的裤子预测，我哪里做错了？

这是我第一次使用NN进行时间序列预测，我试图按小时预测下周或下个月的到达时间，但结果却很差， 进口

%matplotlib inline
import pandas as pd 
from datetime import datetime
from dateutil.relativedelta import relativedelta
import numpy as np
import matplotlib.pyplot as plt
from sklearn.preprocessing import MinMaxScaler
from tensorflow.keras.preprocessing.sequence import TimeseriesGenerator # Generates batches for sequence data
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense,LSTM,Dropout
from tensorflow.keras.callbacks import EarlyStopping  
**Data processing**
def parser(x):
    return datetime.strptime(x,'%Y-%m-%d %H:%M:%S')
cols= ['datetime','arrivals']#,'departures','occupancy']
df = pd.read_csv('/Users/mikesmith/Downloads/last3wkOcc.csv',usecols=cols,parse_dates=['datetime'],index_col=['datetime'],date_parser=parser)
print(df.head())
print(df.tail())
print(df.dtypes)

                    arrivals
datetime                     
2020-07-22 11:00:00  7.0     
2020-07-22 12:00:00  9.0     
2020-07-22 13:00:00  4.0     
2020-07-22 14:00:00  6.0     
2020-07-22 15:00:00  15.0    
                     arrivals
datetime                     
2020-08-12 07:00:00  2.0     
2020-08-12 08:00:00  4.0     
2020-08-12 09:00:00  2.0     
2020-08-12 10:00:00  4.0     
2020-08-12 11:00:00  0.0     
arrivals    float64
dtype: object

print('Number of 60min Buckets ',len(df))# 505
test_pecent = 0.3 # 30 percent of data
len(df)*test_pecent 
test_point = np.round(len(df)*test_pecent) 
test_index =int(len(df) - test_point)
print('test index',test_index)
print('test date start',df.index[test_index])
train = df.iloc[:test_index]
test = df.iloc[test_index:]

Number of 60min Buckets  505
test index 353
test date start 2020-08-06 04:00:00

缩放数据

scaler = MinMaxScaler()
scaler.fit(train)
MinMaxScaler(copy=True,feature_range=(0,1))
scaled_train = scaler.transform(train)
scaled_test = scaler.transform(test)

LSTM模型

early_stop = EarlyStopping(monitor='val_loss',patience=2)
length = 84
batch_size = 1
n_features = 1

generator = TimeseriesGenerator(scaled_train,scaled_train,length=length,batch_size=batch_size)

validation_generator = TimeseriesGenerator(scaled_test,scaled_test,batch_size=batch_size)

model = Sequential()

model.add(LSTM(512,activation = 'relu',input_shape=(length,n_features)))

model.add(Dropout(0.2))

model.add(Dense(1))

model.compile(optimizer = 'adam',loss='mse',metrics=['accuracy'])

model.fit(generator,epochs=20,validation_data=validation_generator,callbacks=[early_stop])

model.summary()

WARNING:tensorflow:sample_weight modes were coerced from
  ...
    to  
  ['...']
WARNING:tensorflow:sample_weight modes were coerced from
  ...
    to  
  ['...']
Train for 269 steps,validate for 68 steps
Epoch 1/20
269/269 [==============================] - 42s 157ms/step - loss: 0.0175 - accuracy: 0.2342 - val_loss: 0.0032 - val_accuracy: 0.3235
Epoch 2/20
269/269 [==============================] - 34s 125ms/step - loss: 0.0149 - accuracy: 0.2342 - val_loss: 0.0051 - val_accuracy: 0.3235
Epoch 3/20
269/269 [==============================] - 37s 137ms/step - loss: 0.0135 - accuracy: 0.2342 - val_loss: 0.0057 - val_accuracy: 0.3235
Model: "sequential_22"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
lstm_22 (LSTM)               (None,512)               1052672   
_________________________________________________________________
dropout_3 (Dropout)          (None,512)               0         
_________________________________________________________________
dense_21 (Dense)             (None,1)                 513       
=================================================================
Total params: 1,053,185
Trainable params: 1,185
Non-trainable params: 0
_________________________________________________________________

预测

forecast = []

first_eval_batch = scaled_train[-length:]

current_batch = first_eval_batch.reshape((1,length,n_features))


for i in range(length):
  current_pred = model.predict(current_batch)[0]

  forecast.append(current_pred)

  current_batch = np.append(current_batch[:,1:,:],[[current_pred]],axis = 1)

绘制预测值

df_forecast = pd.DataFrame(scaler.inverse_transform(forecast),index=df[-length:].index,columns = ['predictions'])
df_test = pd.concat ([df,df_forecast],axis =1)
print(df_test.head())
print('---------------------')
print(df_test.tail(7))

print('number of geusses',len((df_test[df_test['predictions'].notna()])))

                     arrivals  predictions
datetime                                  
2020-07-22 11:00:00  7.0      NaN         
2020-07-22 12:00:00  9.0      NaN         
2020-07-22 13:00:00  4.0      NaN         
2020-07-22 14:00:00  6.0      NaN         
2020-07-22 15:00:00  15.0     NaN         
---------------------
                     arrivals  predictions
datetime                                  
2020-08-12 05:00:00  2.0       3.716657   
2020-08-12 06:00:00  4.0       3.715892   
2020-08-12 07:00:00  2.0       3.714540   
2020-08-12 08:00:00  4.0       3.712937   
2020-08-12 09:00:00  2.0       3.711422   
2020-08-12 10:00:00  4.0       3.710272   
2020-08-12 11:00:00  0.0       3.709655   
number of geusses 84

绘制结果

plt.plot(df_test.index,df_test['arrivals'])
plt.plot(df_test.index,df_test['predictions'])
plt.show()

因此，您可以看到结果不是很准确，不确定我犯了什么错误，我尝试了长度和批次大小，任何帮助或评论都非常感谢