如何将二维Pandas时间序列数组适配到Keras LSTM层？

Let's break down why you're getting that ValueError: Input 0 is incompatible with layer lstm_1: expected ndim=3, found ndim=4 error and how to fix it:

Root Cause #

LSTM layers in Keras expect input data in a 3D format: (number_of_samples, time_steps, number_of_features). Here's what went wrong in your code:

• Your X_train after scaling is 2D ((3180, 8)), but you didn't reshape it to 3D before feeding into the LSTM.
• You passed three values to input_shape ((1, len(X_train), x.shape[1])), which tells Keras to expect a 4D input (batch size + those three dimensions). But LSTM only accepts 3D input, hence the mismatch.

Step-by-Step Fix #

1. Reshape your training data to 3D
   After scaling, convert your 2D array into the 3D format LSTM needs. If you want to treat each row as a single time step (sequence length of 1), use:

   sc = StandardScaler()
   X_train = sc.fit_transform(X_train)
   print(X_train.shape)  # Output: (3180, 8)
   
   # Reshape to 3D: (samples, time_steps, features)
   X_train = X_train.reshape((X_train.shape[0], 1, X_train.shape[1]))
   print(X_train.shape)  # Now outputs: (3180, 1, 8)
   

2. Correct the LSTM input_shape
   The input_shape parameter should only specify (time_steps, features) (omit the sample/batch dimension, which Keras handles automatically). Update your LSTM layer like this:

   classifier = Sequential()
   # input_shape is (time_steps, features) = (1, 8)
   classifier.add(LSTM(units=128, input_shape=(1, X_train.shape[2])))
   

Bonus: Using Longer Sequences (If Needed) #

If you're working with time series and want to use past N time steps to predict future values, you'll need to create sequences first. For example, using sequences of length 10:

import numpy as np

def create_sequences(data, seq_length):
    sequences = []
    for i in range(len(data) - seq_length):
        # Grab consecutive seq_length rows as one sequence
        sequences.append(data[i:i+seq_length])
    return np.array(sequences)

seq_length = 10
X_train = create_sequences(X_train, seq_length)
print(X_train.shape)  # Output: (3170, 10, 8)

# Now update input_shape to match the sequence length
classifier.add(LSTM(units=128, input_shape=(seq_length, X_train.shape[2])))

内容的提问来源于stack exchange，提问作者Manthan