Concatenate prediction values LSTM Keras

Question

I don't know if I'm wrong or if it makes sense but, my idea is to use a single prediction and based on that one, predict the future values (I'm using a LSTM model on Keras). What I'm trying to do is:

1) Get the first X known values (initial_values)

2) Predict using initial_values and get (initial_prediction).

3) Delete the first element of initial_values and append initial_prediction

4) Repeat from step 2) X times.

My code look like this:

predictions = []
y_test_concat = []
num_steps_before = 30

# Step 1

# X_test_scaled_shape: (192, 1)
y_test_concat.append(X_test_scaled[:num_steps_before,:]) 
y_test_concat = np.array(y_test_concat)
y_test_concat.reshape(y_test_concat.shape[0],y_test_concat.shape[1],1)

# Step 2

simulated_predictions.append(model.predict(y_test_concat))

# Step 3 (where I get stucked)

num_steps_to_predict = 10
for i in range(1,num_steps_to_predict):
  ...

So during the next iterations the array should like this:

[initial_value2,initial_value3,...initial_value30, initial_prediction]
[initial_value3,initial_value4,...initial_prediction, initial_prediction2]
...
[initial_value20,initial_value21,...initial_predictionX, initial_predictionY]

Any ideas? I was wondering if there's something already implemented in Keras to do such thing using LSTM.

Can you explain your problem in more detail? In most cases, your input to an LSTM will be a time series and it will keep its own hidden state over that time series. It does seem a little strange that you are using a model's own prediction as its input. — Luke DeLuccia
The idea is based on one prediction, with some initial values, predict the future X next values. What I always see are predictions but using already known data. So you use [x0...x20] and you get [p1] then you use [x1...x21] and get [p2], my idea is use the predicted value (p1) as an input for my input array, so it would be something like [x1...p1] and the next one... [x2...p1, p2] so I'm adding the predictions as part of the next inputs. Does it make sense? Am I missing something with LSTM? — lucaswerner

lucaswerner lucaswerner · Accepted Answer · 2019-02-20T23:32:35

Thanks to @lukedeluccia for his answer, I came up with the solution:

num_steps_before = 30
initial_values = np.array([X_test_scaled[:num_steps_before,:]])
predictions = initial_values

for i in range(0,num_steps_before):
  prediction = model.predict(predictions)
  predictions = np.append(predictions,[prediction])
  predictions = np.delete(predictions,0)
  predictions = predictions.reshape((1,predictions.shape[0],1))

Concatenate prediction values LSTM Keras

1 Answers