如何改进LSTM模型以提升时间序列温度预测精度？

核心问题识别与修改建议 #

1. 任务类型匹配错误（最致命）

温度预测属于回归任务，但当前代码误用了二分类任务的损失函数和评价指标：

• 替换binary_crossentropy为回归任务常用的mse（均方误差）或mae（平均绝对误差）
• 替换accuracy评价指标为mae、mse或mape（平均绝对百分比误差）
• 将模型Checkpoint监控的val_accuracy改为val_loss，确保保存最优的回归模型权重

2. LSTM输入格式错误

LSTM要求输入维度为[样本数, 时间步长, 特征数]，但当前代码未构造时间序列滑动窗口，直接传入一维数据，完全不符合模型输入要求：

• 需编写函数将原始序列转换为滑动窗口格式，用过去N个时间步的温度数据预测下一个时间步的温度，示例逻辑：

def create_sequences(data, seq_length):
    X, y = [], []
    for i in range(len(data) - seq_length):
        X.append(data[i:i+seq_length])
        y.append(data[i+seq_length])
    return np.array(X), np.array(y)

# 构造训练/验证集序列（示例用过去10个时间步预测下一个）
Xtra_seq, Ytra_seq = create_sequences(Xtra, seq_length=10)
Xval_seq, Yval_seq = create_sequences(Xval, seq_length=10)

3. 重采样参数错误

注释标注要转为2小时数据，但代码中用了resample('48H')（48小时=2天），过度压缩数据会丢失时间序列关键细节，应改为resample('2H').mean()

4. 特征单一化问题

仅用温度自身作为输入，无法捕捉时间序列的周期性（季节、月份）和趋势特征：

• 从时间索引中提取衍生特征，示例代码：

df['month'] = df.index.month
df['quarter'] = df.index.quarter
input_columns = ['warm well temperature (°C)', 'month', 'quarter']
X = df[input_columns]

5. 模型与训练参数优化

• LSTM循环层激活函数错误：默认tanh更适合循环网络，替换relu为tanh
• 学习率过高：Adam(learning_rate=0.01)易导致训练震荡，建议降低到0.001或0.0001
• 添加早停机制防止过拟合：

from tensorflow.keras.callbacks import EarlyStopping
early_stop = EarlyStopping(monitor='val_loss', patience=20, restore_best_weights=True)
# 训练时加入callbacks=[modelCheckpoint, early_stop]

• 调整模型容量：当前H=150可能过大，可尝试减小到32/64，或添加Dropout层抑制过拟合：

x = LSTM(H, activation='tanh', return_sequences=False)(i)
x = Dropout(0.2)(x) # 添加Dropout层

原代码 #

#### Link to Google Drive and load modules

# generic modules
import datetime, os
import itertools
import time
import pickle

# basic data science modules
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
%matplotlib inline
import seaborn as sns

# keras
from tensorflow.keras.layers import Input, Dense, Dropout
from tensorflow.keras.models import Model
from tensorflow.keras.optimizers import Adam

# sklearn helper functions
from sklearn.preprocessing import MinMaxScaler
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score, f1_score, classification_report, confusion_matrix

# Import LSTM-related modules
from tensorflow.keras.layers import LSTM


df = pd.read_csv('Master CSV v1.csv',index_col='Date/Time',parse_dates=True)
# Convert 5-minute data to 48-hour data

# Resample the dataframe to 2-hour frequency and take the mean value within each 2-hour interval
df = df.resample('48H').mean()

df = pd.DataFrame(df, index=df.index)

#### inputs and outpts
Y = df['warm well temperature (°C)']

input_columns = ['warm well temperature (°C)']
X = df[input_columns]

# Split the dataset into training and validation sets
Xtra, Xval, Ytra, Yval = train_test_split(X, Y, test_size=0.20, shuffle=False)

# Convert DataFrame to NumPy array
Xtra = Xtra.values
Xval = Xval.values

# Rescale input and output datasets
Sx = MinMaxScaler(feature_range=(0, 1))
Sy = MinMaxScaler(feature_range=(0, 1))

Xtra = Sx.fit_transform(Xtra.reshape(Xtra.shape[0], -1)) 
Xval = Sx.transform(Xval.reshape(Xval.shape[0], -1)) 

Ytra = Sy.fit_transform(Ytra.values.reshape(-1, 1)) 
Yval = Sy.transform(Yval.values.reshape(-1, 1))

# function for LSTM model

def LSTMmodel(L, D, H):
    i = Input(shape=(L, D), name='input_layer')
    x = LSTM(H, activation='relu', name='recurrent_layer', return_sequences=False)(i)
    x = Dense(1, name='output_layer')(x)
    model = Model(i, x, name='LSTM')
    return model

# Create an instance of the LSTM model
lstm_model = LSTMmodel(L=130, D=1, H=150)

# modelCheckpoint
from tensorflow.keras.callbacks import ModelCheckpoint
modelCheckpoint = ModelCheckpoint('LSTM_model_weights v1.hdf5', save_best_only=True, monitor='val_accuracy',mode='auto', save_weights_only=True)

# Compile and train the LSTM model using the fit function
lstm_model.compile(loss='binary_crossentropy', optimizer=Adam(learning_rate=0.01), metrics=['accuracy'])
r = lstm_model.fit(Xtra, Ytra, epochs=500, validation_data=(Xval, Yval), verbose=0, batch_size=64, callbacks=[modelCheckpoint])

# load best model weights
lstm_model.load_weights('LSTM_model_weights v1.hdf5')

# Function for plotting training history
def plot_training_history(r, figsize=(10, 3)):
    f, axes = plt.subplots(1, 2, figsize=figsize)

    # Loss
    axes[0].plot(r.history['loss'], label='Training Loss')
    axes[0].plot(r.history['val_loss'], label='Validation Loss')
    axes[0].set_title('Loss Trajectories')
    axes[0].set_xlabel('Epochs')
    axes[0].set_ylabel('Loss')
    axes[0].legend()

    # Accuracy
    axes[1].plot(r.history['accuracy'], label='Training Accuracy')
    axes[1].plot(r.history['val_accuracy'], label='Validation Accuracy')
    axes[1].set_title('Accuracy Trajectories')
    axes[1].set_xlabel('Epochs')
    axes[1].set_ylabel('Accuracy')
    axes[1].legend()

    # Adjust plot
    plt.tight_layout()
    plt.show()
    return f, axes

# Call the modified function
plot_training_history(r, figsize=(10, 3))

内容的提问来源于stack exchange，提问作者amir