Hi @saifkhanengr,
Looking at the distribution plots, you can see how the data in many features is heavily skewed. Looking at the Boxplots, you can also see that there are a lot of outliers.
All of this affects heavily the machine learning models and the deep learning models. This needs a lot of work, which i can’t possibly detail in here.
I will give you some suggestions. You would need to try to normalize the distributions wherever possible. Some distribution have 0 values and then a normal distribution, bimodal distributions, etc… which you need to deal with.
Remove the outliers (use quantiles) which are definitely affecting the model performance and probably driving the bias in the outcome.
I hope this helps.
Regards,
Samuel
2 Likes
Thank you for your insights, Samuel…
1 Like
Hello Everyone! I hope you all are doing well.
After giving more time to hyperparameters tunning, this is what I got:
One Conv1D with filters=284, kernel_size=3, one LSTM with 284 units, relu, dropout with 0.5, and one dense layer with sigmoid. Loss is MeanSquaredLogarithmicError, rather than mse.
However, my intended Professor said that this is still the wrong approach to use AI in the petroleum industry. He said you must have to include spatiotemporal data, not only one of them, for commercial purposes. And he is absolutely right.
1 Like
Thanks for posting, @saifkhanengr!
Looks like a real exciting project! Congrats 
How long is the prediction horizon (in your plot)?
I always made good experience with adding a (naive) prediction (e.g. derived by a linear regression model or even a constant prediction) as a benchmark to evaluate how well the model of interest performs in that prediction horizon in comparison, see also this thread.
Did you try that already?
Keep it up, @saifkhanengr! 
Best regards
Christian
1 Like
Thank you, Christian!
It’s 14 months for the test set.
Yes, I set a simple RNN as a benchmark (baseline model) and then trained many different models and compared them with that baseline. The best one is the LSTNet (Long- and Short-term Time-series network ), a combination of LSTM and Convolution.
