本文为为🔗365天深度学习训练营内部文章
原作者：K同学啊
一 前期准备数据
1.导入数据
python
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
from sklearn.preprocessing import MinMaxScaler
from keras.layers import Dense,LSTM,Bidirectional
from keras import Input
from keras.optimizers import Adam
from keras.models import Sequential
from sklearn.metrics import mean_squared_error,r2_score
import warnings
warnings.filterwarnings('ignore')
df = pd.read_csv('woodpine2.csv')
print(df)
2.数据可视化
python
'''
可视化
'''
fig,ax = plt.subplots(1,3,constrained_layout=True,figsize=(14,3))
sns.lineplot(data=df['Tem1'],ax=ax[0])
sns.lineplot(data=df['CO 1'],ax=ax[1])
sns.lineplot(data=df['Soot 1'],ax=ax[2])
plt.show()
df = df.iloc[:,1:]
二 特征工程
1.构建特征变量和目标变量
python
'''
用于从一个 Pandas DataFrame 中提取时间序列数据，生成输入（x）和输出（y）数组，适合用于训练机器学习模型
'''
# 取前八个时间段的Tem、CO 1、Soot 1的数据为X，第九个数据为y
width_x = 8
width_y = 1
x = []
y = []
in_start = 0
for _,_ in df.iterrows():
in_end = in_start + width_x
out_end = in_end + width_y
if out_end < len(df):
X_ = np.array(df.iloc[in_start:in_end,])
# 将 X_ 重塑为一维数组，长度为 len(X_) * 3。假设原始数据有多列，这样做的目的是将多列数据展开为一个长数组
X_ = X_.reshape(len(X_)*3)
# 0代表只提取第一列
y_ = np.array(df.iloc[in_end:out_end,0])
x.append(X_)
y.append(y_)
in_start += 1
x = np.array(x)
y = np.array(y)
print(x.shape,y.shape)(5939, 24) (5939, 1)2.归一化
python
'''
归一化
'''
sc = MinMaxScaler(feature_range=(0,1))
x_scaled = sc.fit_transform(x)
# 许多机器学习模型，尤其是 RNN 和 CNN，期望输入数据具有特定的形状。
# 例如，对于 RNN，输入通常是三维的，形状为 (样本数, 时间步数, 特征数)
x_scaled = x_scaled.reshape(len(x_scaled),width_x,3)3.划分训练集和测试集
python
'''
划分训练集
'''
X_train = np.array(x_scaled[:5000]).astype('float64')
y_train = np.array(y[:5000]).astype('float64')
X_test = np.array(x_scaled[5000:]).astype('float64')
y_test = np.array(y[5000:]).astype('float64')
print(X_train.shape,X_test.shape)(5000, 8, 3) (939, 8, 3)三 构建模型
python
'''
构建模型
'''
# 多层LSTM
model_lstm = Sequential()
# units=64: LSTM 单元的数量，决定了模型的学习能力。64 表示该层有 64 个隐藏单元。
# activation='relu': 使用 ReLU 激活函数，帮助引入非线性。虽然 LSTM 通常使用 tanh，但在这里可能是为了与后续层兼容。
# return_sequences=True: 该参数指示 LSTM 层返回每个时间步的输出，而不仅仅是最后一个时间步的输出。这对于后续 LSTM 层是必要的。
# input_shape=(X_train.shape[1], 3): 定义输入数据的形状，其中 X_train.shape[1] 是时间步数（width_x），3 是特征数。
model_lstm.add(LSTM(units=64,activation='relu',return_sequences=True,
input_shape=(X_train.shape[1],3)))
# 由于没有设置 return_sequences=True，该层将只返回最后一个时间步的输出，适用于输出层
model_lstm.add(LSTM(units=64,activation='relu'))
# Dense: 添加全连接层。
# width_y: 输出层的单元数，通常与预测的目标变量数量相同。此处表示模型输出的维度。
model_lstm.add(Dense(width_y))Model: "sequential_1"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
lstm_2 (LSTM) (None, 8, 64) 17408
lstm_3 (LSTM) (None, 64) 33024
dense_1 (Dense) (None, 1) 65
=================================================================
Total params: 50,497
Trainable params: 50,497
Non-trainable params: 0
_________________________________________________________________四 训练模型
python
'''
训练
'''
# 只观测loss数值，不观测准确率，所以删去metrics
model_lstm.compile(optimizer=Adam(1e-3),loss='mean_squared_error')
history_lstm = model_lstm.fit(X_train,y_train,
batch_size=64,
epochs=40,
validation_data=(X_test,y_test),
validation_freq=1)Epoch 1/40
79/79 [==============================] - 2s 10ms/step - loss: 11291.5254 - val_loss: 4030.8381
Epoch 2/40
79/79 [==============================] - 1s 7ms/step - loss: 105.1108 - val_loss: 621.8700
Epoch 3/40
79/79 [==============================] - 1s 7ms/step - loss: 11.1059 - val_loss: 243.1555
Epoch 4/40
79/79 [==============================] - 1s 7ms/step - loss: 7.8829 - val_loss: 226.2837
Epoch 5/40
79/79 [==============================] - 1s 7ms/step - loss: 7.6028 - val_loss: 244.8176
Epoch 6/40
79/79 [==============================] - 1s 7ms/step - loss: 7.7717 - val_loss: 233.2066
Epoch 7/40
79/79 [==============================] - 1s 7ms/step - loss: 7.5230 - val_loss: 223.5901
Epoch 8/40
79/79 [==============================] - 1s 8ms/step - loss: 7.6082 - val_loss: 186.9205
Epoch 9/40
79/79 [==============================] - 1s 8ms/step - loss: 7.6405 - val_loss: 228.2400
Epoch 10/40
79/79 [==============================] - 1s 7ms/step - loss: 7.8450 - val_loss: 189.5714
Epoch 11/40
79/79 [==============================] - 1s 7ms/step - loss: 7.5957 - val_loss: 225.9814
Epoch 12/40
79/79 [==============================] - 1s 8ms/step - loss: 7.8151 - val_loss: 190.3151
Epoch 13/40
79/79 [==============================] - 1s 7ms/step - loss: 7.5514 - val_loss: 171.3298
Epoch 14/40
79/79 [==============================] - 1s 7ms/step - loss: 7.3661 - val_loss: 213.4320
Epoch 15/40
79/79 [==============================] - 1s 8ms/step - loss: 7.5439 - val_loss: 184.0517
Epoch 16/40
79/79 [==============================] - 1s 8ms/step - loss: 7.5511 - val_loss: 171.2384
Epoch 17/40
79/79 [==============================] - 1s 7ms/step - loss: 8.8873 - val_loss: 234.2241
Epoch 18/40
79/79 [==============================] - 1s 7ms/step - loss: 7.4158 - val_loss: 150.4856
Epoch 19/40
79/79 [==============================] - 1s 7ms/step - loss: 7.5648 - val_loss: 181.1954
Epoch 20/40
79/79 [==============================] - 1s 7ms/step - loss: 7.4036 - val_loss: 142.2856
Epoch 21/40
79/79 [==============================] - 1s 7ms/step - loss: 9.1357 - val_loss: 234.9497
Epoch 22/40
79/79 [==============================] - 1s 7ms/step - loss: 8.1056 - val_loss: 161.9010
Epoch 23/40
79/79 [==============================] - 1s 8ms/step - loss: 8.3753 - val_loss: 194.4306
Epoch 24/40
79/79 [==============================] - 1s 8ms/step - loss: 7.6563 - val_loss: 130.8293
Epoch 25/40
79/79 [==============================] - 1s 8ms/step - loss: 7.2591 - val_loss: 132.0350
Epoch 26/40
79/79 [==============================] - 1s 7ms/step - loss: 7.7230 - val_loss: 160.6877
Epoch 27/40
79/79 [==============================] - 1s 7ms/step - loss: 7.5739 - val_loss: 174.5122
Epoch 28/40
79/79 [==============================] - 1s 7ms/step - loss: 6.6420 - val_loss: 134.3670
Epoch 29/40
79/79 [==============================] - 1s 7ms/step - loss: 8.0128 - val_loss: 186.9790
Epoch 30/40
79/79 [==============================] - 1s 7ms/step - loss: 7.7388 - val_loss: 175.0219
Epoch 31/40
79/79 [==============================] - 1s 8ms/step - loss: 7.2826 - val_loss: 162.1556
Epoch 32/40
79/79 [==============================] - 1s 7ms/step - loss: 7.5271 - val_loss: 179.4215
Epoch 33/40
79/79 [==============================] - 1s 7ms/step - loss: 6.5252 - val_loss: 104.9447
Epoch 34/40
79/79 [==============================] - 1s 7ms/step - loss: 6.6822 - val_loss: 250.2031
Epoch 35/40
79/79 [==============================] - 1s 7ms/step - loss: 7.0211 - val_loss: 95.4579
Epoch 36/40
79/79 [==============================] - 1s 7ms/step - loss: 8.2577 - val_loss: 116.2146
Epoch 37/40
79/79 [==============================] - 1s 7ms/step - loss: 7.5878 - val_loss: 243.1991
Epoch 38/40
79/79 [==============================] - 1s 7ms/step - loss: 11.2559 - val_loss: 390.6861
Epoch 39/40
79/79 [==============================] - 1s 7ms/step - loss: 7.0651 - val_loss: 168.2283
Epoch 40/40
79/79 [==============================] - 1s 7ms/step - loss: 7.6015 - val_loss: 223.1200五 评估模型
python
'''
评估
'''
# loss图
plt.rcParams['font.sans-serif'] = ['SimHei']
plt.rcParams['axes.unicode_minus'] = False
plt.figure(figsize=(5,3))
plt.plot(history_lstm.history['loss'],label='LSTM Training Loss')
plt.plot(history_lstm.history['val_loss'],label='LSTM Testing Loss')
plt.title('Training and Testing Loss')
plt.legend()
plt.show()
python
# 调用模型预测
pred_y = model_lstm.predict(X_test)
y_test_one = [i for i in y_test]
pred_y_one = [i[0] for i in pred_y]
plt.figure(figsize=(8,6))
plt.plot(y_test_one[:1000],color='red',label='真实值')
plt.plot(pred_y_one[:1000],color='blue',label='预测值')
plt.legend()
plt.show()
python
RMSE_lstm = mean_squared_error(pred_y,y_test) ** 0.5
R2_lstm = r2_score(pred_y,y_test)
print('均方差误差：',RMSE_lstm)
print('R2:',R2_lstm)均方差误差： 14.93720339824608
R2: 0.5779344291884085总结：
捕捉时间序列中的长期依赖性
处理不规则的时间间隔和噪声
建模复杂的非线性关系
动态调整记忆和遗忘
适应性强