基于RNN(LSTM)的时间序列数据集增强异常问题求助

问题背景 #

使用LSTM模型增强IoT传感器生成的1小时间隔时间序列数据，目标是将20/67/280天的训练数据集扩展X天合成数据。但生成的合成数据后期全部变为相同值，即使将训练轮次从100增至200仍未解决。原始数据无缺失，呈线性特征。

数据样本 #

原始数据 #

timestamp,value
2013-07-04 00:00:00,69.88083514
2013-07-04 01:00:00,71.22022706
2013-07-04 02:00:00,70.87780496
2013-07-04 03:00:00,68.95939994
2013-07-04 04:00:00,69.28355102
2013-07-04 05:00:00,70.06096581
2013-07-04 06:00:00,69.27976479
2013-07-04 07:00:00,69.36960846
....

合成数据异常表现 #

....
2014-03-02 02:00:00,65.00628144
2014-03-02 03:00:00,65.10276279
2014-03-03 09:00:00,64.73752596
....
2014-03-03 09:00:00,64.90076446533203
2014-03-03 10:00:00,65.06415557861328
2014-03-03 11:00:00,65.22787475585938
2014-03-03 12:00:00,65.39212799072266
2014-03-03 13:00:00,65.5571060180664
2014-03-03 14:00:00,65.7229995727539
2014-03-03 15:00:00,65.8899917602539
....
2014-03-09 09:00:00,85.6402587890625
2014-03-09 10:00:00,85.64026641845703
2014-03-09 11:00:00,85.64026641845703
2014-03-09 12:00:00,85.64026641845703
2014-03-09 13:00:00,85.64026641845703
2014-03-09 14:00:00,85.64026641845703
2014-03-09 15:00:00,85.64026641845703
2014-03-09 16:00:00,85.64026641845703

现有实现代码 #

import pandas as pd
import numpy as np
from keras.layers import LSTM, Dense
from keras.models import Sequential

def augment_timeseries(df, num_days):

    # Create a copy of the timestamp column
    timestamp_col = df['timestamp'].copy()

    # Convert timestamp column to datetime format
    df['timestamp'] = pd.to_datetime(df['timestamp'])

    # Set timestamp column as the index
    df.set_index('timestamp', inplace=True)
    
    # Get the frequency of the timestamps in the original dataset
    freq = 'H'
    
    # Create a range of timestamps with the same frequency as the original dataset
    end_time = df.index[-1] + pd.DateOffset(days=num_days)
    timestamps = pd.date_range(start=df.index[-1], end=end_time, freq=freq)

    # Reshape data for input into RNN model
    X_train = df.values.reshape((-1, 1, 1))

    # Create the RNN model
    model = Sequential()
    model.add(LSTM(50, input_shape=(1, 1), return_sequences=True))
    model.add(Dense(1))
    model.compile(loss='mean_squared_error', optimizer='adam')

    # Fit the model on the data
    model.fit(X_train, X_train, epochs=100, batch_size=1, verbose=1)

    # Use the model to make predictions for new timestamps
    for timestamp in timestamps:
        X_test = np.array([[df.values[-1]]])
        prediction = model.predict(X_test)

        df.loc[len(df)] = [prediction[0][0][0]]

        timestamp_df = pd.DataFrame([timestamp])
        timestamp_col = pd.concat([timestamp_col, timestamp_df], ignore_index=True)
        
        model.reset_states()

    # Add the timestamp column back to the original dataframe
    df.insert(0, 'timestamp', timestamp_col.values)

    return df

解决思路与技术修正 #

• 重构训练数据输入逻辑：当前代码仅用单个值预测下一个值，且每次预测后重置模型状态，完全浪费了LSTM的序列记忆能力。应使用滑动窗口构建训练样本，比如用过去24小时（对应日周期）的数据作为输入，预测下一小时的值，让模型学习序列的趋势与周期特征。示例：将X_train构造成(样本数, 时间步长, 特征数)的格式，时间步长设为24。
• 调整模型结构：当前LSTM设置return_sequences=True但仅接Dense层，输出为序列而非单值。单步预测场景下，应将return_sequences设为False，或在最后一层LSTM后接Dense输出单值。同时可加入Dropout层（如model.add(Dropout(0.2))）防止过拟合。
• 添加数据归一化：LSTM对数据尺度敏感，建议用MinMaxScaler将数据缩放到[0,1]区间，预测后再反缩放还原，避免数值漂移导致的收敛问题。
• 优化预测流程：取消每次预测后的model.reset_states()调用，让模型保持对生成序列的记忆；或采用多步预测方式，一次生成多个时间步的结果，而非单步迭代。
• 调整训练策略：降低Adam优化器的学习率（如learning_rate=0.0001），增大batch_size提升训练稳定性；加入验证集监控过拟合，避免盲目增加训练轮次。
• 适配线性特征：原始数据呈线性，可尝试先对数据做差分处理消除趋势，训练生成差分序列后再还原趋势；或在模型中结合线性层，提升对线性特征的拟合能力。

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