LSTM输入形状不匹配报错，需用30行数据集预测第31个值

Let's break down your error and fix this step by step. The core issue is a mismatch between the input shape you defined for your LSTM and the actual shape of your training data, plus a few other small bugs in the prediction and visualization code.

1. Root Cause of the Shape Mismatch #

Your LSTM's input_shape is set to (X_train.shape[0], 5) which is completely wrong:

• X_train.shape[0] is the number of training samples (20 in your case, since you have 30 rows and use 10 timesteps), not the number of timesteps.
• You only have 1 feature (the pow column), not 5.

The correct input shape for LSTM should be (number_of_timesteps, number_of_features), which for your data is (10, 1).

2. Step-by-Step Fixes #

Let's go through each part of your code and correct the issues:

Fix 1: Correct the LSTM Input Shape #

When defining your first LSTM layer, change the input_shape parameter to match your training data's timesteps and features:

regressior.add(LSTM(units=60, activation='relu', return_sequences=True, input_shape=(X_train.shape[1], X_train.shape[2])))

X_train.shape[1] gives the timesteps (10), X_train.shape[2] gives the number of features (1).

Fix 2: Fix the Prediction Code #

Your current prediction code has multiple issues:

• You're using a single value new = 62 instead of the last 10 timesteps from your data (since your model is trained on 10 timesteps to predict the next value).
• scaler.transform requires a 2D array, not a scalar.
• You have a typo: regressior.predict(input) should be regressior.predict(inputs).

Replace the prediction section with this:

# Get the last 10 values from the scaled training data to predict the 31st value
new_data = data_training[-10:]  # Shape (10,1)
# Reshape to match the model's expected input shape (samples, timesteps, features)
inputs = new_data.reshape(1, 10, 1)
# Predict and inverse transform to get the actual power value
y_pred_scaled = regressior.predict(inputs)
y_pred = scaler.inverse_transform(y_pred_scaled)
print(f"Predicted 31st power value: {y_pred[0][0]}")

Fix 3: Fix Visualization Code #

You can't plot the 3D X_train directly. Instead, plot the original training data alongside the model's predictions on the training set (to see how well it fits):

# Get predictions on training data to visualize
train_predict_scaled = regressior.predict(X_train)
train_predict = scaler.inverse_transform(train_predict_scaled)
y_train_actual = scaler.inverse_transform(y_train.reshape(-1,1))

plt.figure(figsize=(14,5))
plt.plot(y_train_actual, color='red', label='Real Power Data')
plt.plot(train_predict, color='blue', label='Predicted Power Data')
plt.title('Power Prediction (Training Set)')
plt.xlabel('Time Step')
plt.ylabel('Power')
plt.legend()
plt.show()

Fix 4: Optional: Adjust Batch Size (Since Dataset is Small) #

Your dataset only has 20 training samples, so a batch size of 32 is larger than the number of samples. Reduce it to 8 or 16 to avoid warnings:

regressior.fit(X_train, y_train, epochs=50, batch_size=8)

3. Full Corrected Code #

Here's the complete fixed code:

import numpy as np
import matplotlib.pyplot as plt
import pandas as pd
from sklearn.preprocessing import MinMaxScaler
from tensorflow.keras import Sequential
from tensorflow.keras.layers import Dense, LSTM, Dropout

# Load and preprocess data
data_training = pd.read_csv('G:\\Agrima\\Obj_Rec\\power.csv')
# Extract the pow column (assuming it's the 3rd column, index 2)
data_training = data_training.iloc[:, 2:3].values

scaler = MinMaxScaler()
data_training = scaler.fit_transform(data_training)
print(data_training[0:10])

# Prepare training sequences (10 timesteps to predict next value)
X_train = []
y_train = []
for i in range(10, data_training.shape[0]):
    X_train.append(data_training[i-10:i])
    y_train.append(data_training[i, 0])
X_train, y_train = np.array(X_train), np.array(y_train)
print(f"X_train shape: {X_train.shape}")  # Should be (20, 10, 1)
print(f"y_train shape: {y_train.shape}")  # Should be (20,)

# Build LSTM model with correct input shape
regressor = Sequential()
regressor.add(LSTM(units=60, activation='relu', return_sequences=True, input_shape=(X_train.shape[1], X_train.shape[2])))
regressor.add(Dropout(0.2))
regressor.add(LSTM(units=60, activation='relu', return_sequences=True))
regressor.add(Dropout(0.2))
regressor.add(LSTM(units=80, activation='relu', return_sequences=True))
regressor.add(Dropout(0.2))
regressor.add(LSTM(units=120, activation='relu'))
regressor.add(Dropout(0.2))
regressor.add(Dense(units=1))

print(regressor.summary())
regressor.compile(optimizer='adam', loss='mean_squared_error')

# Train with appropriate batch size
regressor.fit(X_train, y_train, epochs=50, batch_size=8)
regressor.save_weights('my_model_weights.h5')

# Predict the 31st value
new_data = data_training[-10:]
inputs = new_data.reshape(1, 10, 1)
y_pred_scaled = regressor.predict(inputs)
y_pred = scaler.inverse_transform(y_pred_scaled)
print(f"Predicted 31st Power Value: {y_pred[0][0]}")

# Visualize training predictions
train_predict_scaled = regressor.predict(X_train)
train_predict = scaler.inverse_transform(train_predict_scaled)
y_train_actual = scaler.inverse_transform(y_train.reshape(-1,1))

plt.figure(figsize=(14,5))
plt.plot(y_train_actual, color='red', label='Real Power Data')
plt.plot(train_predict, color='blue', label='Predicted Power Data')
plt.title('Power Prediction (Training Set)')
plt.xlabel('Time Step')
plt.ylabel('Power')
plt.legend()
plt.show()

Key Notes #

• Since your dataset is very small (only 30 rows), the model might overfit easily. You might want to reduce the number of LSTM layers or units, or add more regularization.
• Always double-check the shape of your data at each step using print(X_train.shape) to catch shape mismatches early.

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