LSTM RNN 同时预测多个时间步长和多个特征

Question

I have a dataset from 4 temperature sensors measuring different places in/around a building:

I'm training a model that takes inputs of shape (96, 4), 96 time steps for the 4 sensors. From this I want to predict 48 points into the future for each of those sensors, shape (48, 4).

So far I've got an implementation working to predict one sensor only. I mostly followed this section from the TensorFlow tutorials .

My train X is shape (6681, 96, 4), train Y is shape (6681, 48) as I have restricted this to one sensor only. If I just change train Y to (6681, 48, 4) when training I of course get ValueError: Dimensions must be equal, but are 48 and 4 for 'loss/dense_loss/sub' (op: 'Sub') with input shapes: [?,48], [?,48,4]. as my model is not expecting this shape.

Where I'm getting stuck is with my LSTM layer's input/output shapes. I just can't figure out how to finish with a shape of (BATCH_SIZE, 48, 4).

Here's my layer setup at the moment:

tf.keras.backend.clear_session()


print("Input shape", x_train_multi.shape[-2:])

multi_step_model = tf.keras.models.Sequential()
multi_step_model.add(tf.keras.layers.LSTM(32,
                                          return_sequences=True,
                                          input_shape=x_train_multi.shape[-2:]))
multi_step_model.add(tf.keras.layers.Dropout(rate=0.5)) # Dropout layer after each LSTM to reduce overfitting.
multi_step_model.add(tf.keras.layers.LSTM(16, activation='relu'))
multi_step_model.add(tf.keras.layers.Dropout(rate=0.5)) 
# The argument to Dense shapes the results to give the number of time steps we want.
# But how do I make it keep 4 features as well?!?
multi_step_model.add(tf.keras.layers.Dense(future_target / STEP))
multi_step_model.compile(optimizer=tf.keras.optimizers.RMSprop(clipvalue=1.0), loss='mae')

# Shape of predictions
for x, y in val_data_multi.take(1):
    print ("Prediction shape", multi_step_model.predict(x).shape)

Some thoughts:

• Am I just missing something or forgetting to set an argument for the output features/dimensions to use?
• Do I need to train separate RNNs for predicting each sensor?

Thanks! :)

Answer 1

I managed to achieve this in the end by using a Dense layer with the number of time steps I desired multiplied by the number of features I was predicting. Then after that, I reshaped this into the output shape I desired.

I'm not sure if this is the optimal method for doing this, but it works okay.

#Experimental code for predicting multiple sensors
import tensorflow.keras.layers as tfl

tf.keras.backend.clear_session()


print("Input shape", x_train_multi.shape[-2:]) 
# Input shape (96, 4)

multi_step_model = tf.keras.Sequential()
multi_step_model.add(tf.keras.layers.LSTM(32, return_sequences=True, input_shape=x_train_multi.shape[-2:]))
multi_step_model.add(tf.keras.layers.Dropout(rate=0.5))
multi_step_model.add(tf.keras.layers.LSTM(16, return_sequences=False, activation='relu'))
multi_step_model.add(tf.keras.layers.Dropout(rate=0.5)) 
print("After LSTMs", multi_step_model.output_shape)  
# After LSTMs (None, 16)
multi_step_model.add(tf.keras.layers.Dense((future_target / STEP) * 4))
print("After Dense Layer", multi_step_model.output_shape) 
#  After Dense Layer (None, 192)
multi_step_model.add(tf.keras.layers.Reshape((int(future_target / STEP), 4)))
print("After Reshape", multi_step_model.output_shape)
# After Reshape (None, 48, 4)


multi_step_model.compile(optimizer=tf.keras.optimizers.RMSprop(clipvalue=1.0), loss='mae')

# Shape of predictions
for x, y in val_data_multi.take(1):
    print ("Prediction shape", multi_step_model.predict(x).shape)
    # Prediction shape (512, 48, 4)