基于GridSearchCV的LSTM模型训练异常：损失高精度低求助

问题背景 #

尝试使用GridSearchCV构建LSTM深度神经网络模型，但训练后模型损失值极高、精度值极低，无法定位问题根源。

代码重现 #

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import tensorflow as tf
from sklearn.preprocessing import MinMaxScaler
from sklearn.metrics import r2_score, median_absolute_error, mean_absolute_error, mean_absolute_percentage_error, mean_squared_error
from keras.layers import Dense,Dropout,LSTM, LayerNormalization
from keras.models import Sequential
from datetime import datetime
# from scikeras.wrappers import KerasClassifier, KerasRegressor
from keras.wrappers.scikit_learn import KerasClassifier
from sklearn.model_selection import GridSearchCV
from sklearn.model_selection import RandomizedSearchCV
from sklearn.model_selection import train_test_split
import seaborn as sb

df = pd.read_excel('izmir_saatlik_tuketim_verisi.xlsx', sheet_name='Veri (2)')
time_index = pd.date_range(start='2017-01-01 00:00', periods=len(df),  freq='H')  
time_index = pd.DatetimeIndex(time_index)
df['time']=time_index
df['hour']= df['time'].dt.hour
df['week']= df['time'].dt.week
df['dayofweek']= df['time'].dt.dayofweek
df['dayofyear']= df['time'].dt.dayofyear
df['month']= df['time'].dt.month
df['Unnamed: 0']=df['time']
df.rename(columns ={'Unnamed: 0':'Time'},inplace=True)
df.drop(['time'], axis=1, inplace=True)
df['Nispi_Nem(%)'].fillna(df['Nispi_Nem(%)'].median(), inplace=True)

X=df.iloc[:,2:9].values
y=df.iloc[:,1].values
X_train, X_test, y_train, y_test=train_test_split(X,y, test_size=0.1, random_state=0)

#---------------------------Normalize of data----------------------
scaler=MinMaxScaler()
X_train=scaler.fit_transform(X_train)
X_test=scaler.fit_transform(X_test)
y_train=scaler.fit_transform(y_train.reshape(-1,1))
y_test=scaler.fit_transform(y_test.reshape(-1,1))

X_train = X_train.reshape((X_train.shape[0], 1, X_train.shape[1]))
X_test = X_test.reshape((X_test.shape[0], 1, X_test.shape[1]))

#-------------------------------Built in model-------------------------
param_grid = {'batch_size' : [50,100,250],
                # 'activation':['tanh','sigmoid','relu'],
                # 'optimizer':['adam','rmsprop','adagrad'],
                'epochs' : [20,50,100],
                # 'units':[16,32,64,128,256]
                }
Dropout_rate=0.5
def create_model(units=16, optimizer='adam'):
    lstm_model = Sequential()
    lstm_model.add(LSTM(units=units,activation='tanh',return_sequences=True, 
                        input_shape=(X_train.shape[1], X_train.shape[2])))
    lstm_model.add(LSTM(units=units,activation='tanh',return_sequences=False))
    lstm_model.add(Dense(1, activation='Softmax'))
    lstm_model.compile(optimizer=optimizer,loss="mae", metrics=["acc"])
    return lstm_model

tf.random.set_seed(111)

#-------------------------Hyperparamaters Optimization------------------------
model = KerasClassifier(build_fn=create_model, verbose=1)

grid = GridSearchCV(estimator=model, param_grid=param_grid, cv = 2, verbose=0)

grid_result=grid.fit(X_train, y_train)

print("Best: %f using %s" % (grid_result.best_score_, grid_result.best_params_))

核心问题与修正方案 #

1. 任务类型混淆（关键错误） #

• 问题：当前是回归任务（预测连续值，如电力消耗），但误用了KerasClassifier（分类任务专用），同时搭配分类指标acc，导致模型优化方向完全错误，精度计算无意义。
• 修正：
  • 取消注释from scikeras.wrappers import KerasRegressor，替换KerasClassifier为KerasRegressor。
  • 编译模型时，将metrics=["acc"]改为回归适用指标，比如metrics=["mse"]，后续评估用r2_score、mean_absolute_error等回归指标。

2. 输出层激活函数错误（关键错误） #

• 问题：回归任务最后一层使用Softmax激活，该函数用于多分类任务的概率归一化，完全不符合回归输出连续值的需求。
• 修正：将输出层改为线性激活（默认即为线性，可省略激活参数）：

  lstm_model.add(Dense(1))  # 等价于 activation='linear'
  

3. 数据归一化流程错误 #

• 问题：测试集使用fit_transform会重新计算归一化参数，导致训练/测试集数据分布不一致，模型无法正确泛化。
• 修正：仅在训练集上拟合归一化器，测试集直接复用训练集的归一化参数：

  # 特征归一化
  scaler_X = MinMaxScaler()
  X_train = scaler_X.fit_transform(X_train)
  X_test = scaler_X.transform(X_test)
  
  # 目标值归一化
  scaler_y = MinMaxScaler()
  y_train = scaler_y.fit_transform(y_train.reshape(-1,1))
  y_test = scaler_y.transform(y_test.reshape(-1,1))
  

4. 时间序列数据拆分错误 #

• 问题：train_test_split随机划分数据集，破坏了时间序列的时序性，LSTM无法学习到时间依赖关系。
• 修正：按时间顺序划分训练/测试集，比如取前90%数据训练，后10%测试：

  split_idx = int(len(df) * 0.9)
  X_train, X_test = X[:split_idx], X[split_idx:]
  y_train, y_test = y[:split_idx], y[split_idx:]
  

5. LSTM输入结构未发挥时序特性 #

• 问题：将数据reshape为(样本数, 1, 特征数)，每个样本仅含1个时间步，LSTM的时序建模能力完全无法体现。
• 修正：构造多步时序输入，比如用过去24小时的数据预测当前值：

  def create_time_sequences(X, y, seq_len):
      X_seq, y_seq = [], []
      for i in range(len(X) - seq_len):
          X_seq.append(X[i:i+seq_len])
          y_seq.append(y[i+seq_len])
      return np.array(X_seq), np.array(y_seq)
  
  # 使用过去24小时的数据作为输入
  seq_length = 24
  X_train_seq, y_train_seq = create_time_sequences(X_train, y_train, seq_length)
  X_test_seq, y_test_seq = create_time_sequences(X_test, y_test, seq_length)
  
  # 此时输入形状为 (样本数, seq_length, 特征数)，符合LSTM时序输入要求
  

6. 未使用定义的Dropout层 #

• 问题：代码定义了Dropout_rate=0.5但未添加Dropout层，后续优化时可能出现过拟合。
• 修正：在LSTM层后添加Dropout层：

  def create_model(units=16, optimizer='adam'):
      lstm_model = Sequential()
      lstm_model.add(LSTM(units=units,activation='tanh',return_sequences=True, 
                          input_shape=(X_train_seq.shape[1], X_train_seq.shape[2])))
      lstm_model.add(Dropout(Dropout_rate))
      lstm_model.add(LSTM(units=units,activation='tanh',return_sequences=False))
      lstm_model.add(Dropout(Dropout_rate))
      lstm_model.add(Dense(1))
      lstm_model.compile(optimizer=optimizer,loss="mae", metrics=["mse"])
      return lstm_model
  

内容的提问来源于stack exchange，提问作者Fatih BERBER