基于LSTM的温度预测：输入数据归一化的最优方案探讨

阿华AIGC实验室

2026-5-22

LSTM温度预测：输入特征的归一化最佳实践

Great question—this is a common pitfall when working with multi-feature time series models like LSTMs, and getting the normalization right can make a huge difference in your model's performance. Let's break this down clearly:

核心结论：必须对每个输入特征单独进行归一化

You should never scale your three input features (temperature, pressure, wind speed) together using Min-Max scaling. Here's why:

每个特征的物理意义和数值范围完全不同：气压可能在950-1050 hPa之间，风速在0-20 m/s，温度在-10-40℃。混在一起缩放会让温度的数值分布被其他特征的极值严重压缩，丢失关键的模式信息。LSTMs rely on consistent feature scales to learn meaningful temporal patterns, so this would cripple your model's ability to focus on temperature-related trends.
你的担忧（输出温度与原始极值不一致）是完全合理的，但这个问题可以通过单独处理目标变量的归一化来解决，和输入特征的缩放完全独立。

具体实现步骤

1. 对每个输入特征使用独立的Min-Max缩放器

For each feature (temperature, pressure, wind speed), fit a separate MinMaxScaler using only your training data (critical to avoid data leakage). This preserves the unique distribution of each feature while bringing them all into the [0,1] range, which works well with LSTM activation functions like tanh.

2. 对输出温度使用单独的缩放器

Since your output is temperature, create a separate MinMaxScaler specifically for the target temperature values. Fit this scaler on the training set's temperature labels, then scale both training and test targets. After making predictions, you'll use this scaler to inverse-transform the predicted values back to the original temperature range.

3. 关键注意事项：禁止用测试集数据拟合缩放器

Always split your data into train/test sets first, then fit all scalers exclusively on the training data. Using test data to fit scalers introduces data leakage, which makes your model's performance look better than it actually is on unseen data.

代码示例（Python + Scikit-learn）

from sklearn.preprocessing import MinMaxScaler
import pandas as pd

# 假设你的数据存储在DataFrame中
data = pd.read_csv("your_time_series_data.csv")

# 第一步：拆分训练集和测试集（按时间顺序拆分，不要随机打乱！）
train_size = int(0.8 * len(data))
train_data = data[:train_size]
test_data = data[train_size:]

# 分离输入特征和目标变量
train_features = train_data[["temperature", "pressure", "wind_speed"]].values
train_target = train_data["temperature"].values.reshape(-1, 1)

test_features = test_data[["temperature", "pressure", "wind_speed"]].values
test_target = test_data["temperature"].values.reshape(-1, 1)

# 初始化单独的缩放器
scaler_temp = MinMaxScaler()
scaler_press = MinMaxScaler()
scaler_wind = MinMaxScaler()
scaler_target = MinMaxScaler()

# 拟合并转换训练集特征
train_features[:, 0] = scaler_temp.fit_transform(train_features[:, 0].reshape(-1, 1)).flatten()
train_features[:, 1] = scaler_press.fit_transform(train_features[:, 1].reshape(-1, 1)).flatten()
train_features[:, 2] = scaler_wind.fit_transform(train_features[:, 2].reshape(-1, 1)).flatten()

# 拟合并转换训练集目标
train_target_scaled = scaler_target.fit_transform(train_target)

# 转换测试集（用训练集拟合的缩放器，不要重新拟合！）
test_features[:, 0] = scaler_temp.transform(test_features[:, 0].reshape(-1, 1)).flatten()
test_features[:, 1] = scaler_press.transform(test_features[:, 1].reshape(-1, 1)).flatten()
test_features[:, 2] = scaler_wind.transform(test_features[:, 2].reshape(-1, 1)).flatten()

# 训练你的LSTM模型...

# 预测后反归一化得到真实温度
predicted_scaled = model.predict(test_features)
predicted_temperature = scaler_target.inverse_transform(predicted_scaled)

替代方案：Z-Score标准化（均值为0，方差为1）

If your features follow a roughly normal distribution, you could use StandardScaler instead of MinMaxScaler. This centers each feature around its mean and scales to unit variance. The same rule applies: use separate scalers for each input feature and the target variable.