LSTM model doesn't predict values higher than certain value (not same value all the time)

Question

First of all thanks for any help!

I want to create a simple LSTM model that predict the value of next minute Household Electric Power Consumption. using this dataset:

https://archive.ics.uci.edu/ml/datasets/individual+household+electric+power+consumption

What I've done so far is:

1) normalize the data and created a "window" that makes my LSTM network look like this:

Series of 30 minutes from t0-t29 with 8 features in each minute, and the network should output the t30 house electric power consumption. Therefore my each sample input is in a shape of [30x8] and output should be [1x1]. so far, so good.

2) I've build a simple LSTM model that looks like this:

model4 = Sequential()
model4.add(LSTM(1,input_shape=(30, 8)))
model4.add(Dense(1))
model4.compile(loss='mean_squared_error', optimizer='adam')

As you can see, a very simple many to one model. I've trained the model for 25 epochs and for some reason the model doesn't seem to predict value higher than 5-5.5 as you can see in this graph:

3) I've tried to:

• normalize the data differently (min max or mean std)
• normalize all data but the column of the power consumption i kept with it original values. None of the above helped.

any suggestions?

Answer 1

Solution:

i changed to property "activation" of the LSTM layer to None. this way the values were not "smashed" by the default "tanh" function.