利用鲸鱼优化算法对LSTM网络的超参数进行优化,LSTM模型如下所示。包含对数据的预处理,时间滑窗函数的编写,LSTM网络模型的定义以及适应度函数的定义,适应度函数即为WOA算法的目标函数。
1.首先导入必须的库。(注意要把GPU禁掉,因为是时间序列数据,量很小,用GPU反而会使得速度变慢)
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import WOA,random
import tensorflow as tf
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Dropout, LSTM
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import MinMaxScaler
from sklearn.metrics import mean_absolute_error,mean_squared_error,mean_absolute_percentage_error,r2_score
import os
os.environ['CUDA_VISIBLE_DEVICES'] = '-1'
读取数据,并且对数据做常规处理,划分训练集和测试集。
#读取数据
df = pd.read_csv('threegorges-water-storage.csv',index_col=0)
#填充空值
df.fillna(value=df.mean(),inplace=True)
#更换列名
df = df.rename(columns={'upstream_water_level': '上游水位', 'downstream_water_level':
'下游水位', 'inflow_rate': '流入速率', 'outflow_rate': '流出速率'})
#选择训练集和测试集
Y=df['流出速率'].values.reshape(-1,1)
X=df.values
#归一化
transfer_X = MinMaxScaler()
transfer_Y = MinMaxScaler()
X_Standar = transfer_X.fit_transform(X)
Y_Standar = transfer_Y.fit_transform(Y)
x_train, x_test, y_train, y_test = train_test_split(X_Standar, Y_Standar,
test_size=0.25, random_state=100,shuffle=False
)
定义时间滑窗函数,这一步比较关键。这一步的目的是方便后续构建训练和测试数据。
#定义时间滑窗函数
def create_time_windows(x, y, window_size, stride):
'''
:param x: 特征值
:param y: 目标值
:param window_size:时间窗格大小
:param stride: 时间步长
:return: 返回增加了时间滑窗的x和y
'''
x_windows = []
y_windows = []
for i in range(0, len(x) - window_size, stride):
x_window = x[i:i+window_size]
y_window = y[i+window_size]
x_windows.append(x_window)
y_windows.append(y_window)
return np.array(x_windows), np.array(y_windows)
构建训练数据和测试数据,时间窗格大小设置为30,时间步设置为1.
window_size = 30 #时间窗格大小
stride = 1 #时间步长
#划分训练集和测试集
x_train, y_train = create_time_windows(x_train, y_train, window_size, stride)
x_test, y_test = create_time_windows(x_test, y_test, window_size, stride)
定义神经网络模型,这种比较简单的模型采用tensorflwo搭建相对来说比pytorh更方便一些。
verbose = 0
##定义神经网络模型
def LSTM_model(units1,units2,Dorpout_rate,lr,batchsize):
model = Sequential()
model.add(LSTM(units1, input_shape=(x_train.shape[1], x_train.shape[2]), return_sequences=True,activation='relu'))
model.add(Dropout(Dorpout_rate))
model.add(LSTM(units2, return_sequences=False,activation='relu'))
model.add(Dense(1, activation='relu'))
model.compile(loss='mse', optimizer=Adam(learning_rate=lr))
history = model.fit(x_train, y_train, epochs=30, batch_size=batchsize, shuffle=False
, validation_data=(x_test, y_test), verbose=verbose)
# 计算适应度函数
y_test_pre = model.predict(x_test)
y_test_pre_invers = transfer_Y.inverse_transform(y_test_pre)
y_test_invers = transfer_Y.inverse_transform(y_test)
return y_test_pre_invers,y_test_invers,history,model
设置适应度函数,同时要设置随机数种子,保证结果可复现,否则优化无意义。
# 适应度函数
def fun(x):
tf.random.set_seed(123) # 设置随机数种子
np.random.seed(123)
random.seed(123)
units1 = int(x[0]) # LSTM1层神经元节点
units2 = int(x[1]) # LSTM层神经元节点
Dorpout_rate = x[2] # Dorpout率
lr = x[3] # 学习率
batchsize = int(x[4]) # 抓取个数
y_test_pre_invers, y_test_invers, history,model = LSTM_model(units1,units2,Dorpout_rate,lr,batchsize)
fit_mse = mean_squared_error(y_test_pre_invers, y_test_invers)
return fit_mse
WOA-LSTM模型的预测结果如下图所示,可以看到预测结果和真实值相比还是比较接近的、
loss曲线如下图,可以看到效果还是比较好的,随着迭代的进行,损失一直在下降。
完整代码和数据加qq1019312261