I am currently developping a forecasting model to forecast the energy production in a thermo-electric center , so i have 31 features and the target, almost 150K records of each 10 mins going back since 2021, all models does so great in the test and validation sets, all metrics are good , but the problem is when i want to predict the next week energy production , they output diverge rapedly (normal production 10-50 MGW/H)(model goes up to +infinity) , i tryied adding date variable and lags also but it still doesnt any better, the energy prooduction column doesnt folow any seationality so it’s almost a random event.
what can i do to make the performance better in the future dates ,(preferably in LSTM or Xgboost)
What method does your model use now?
I generate a sequence of dates that I want to forecast on (a week with 10-minute intervals, for example, so it’s 7 days * 24 hours * 6 dates). I take the last n_steps from the data and loop over the forecast dates (making predictions on the last n_steps, appending the result to a list, rolling it by one step and removing the 1st , and iteratively continuing making prediction using new input sequance)
What type of model are you using - I did not ask how you’re making the predictions.
i am using LSTM
@ABOU_ELKHIR_MOHAMED This is outside of the scope of the courses here, but on the long list of things I’d like to get to, I think FB’s ‘Prophet’ model looks kind of interesting for time series: Prophet | Forecasting at scale.
And then if you are talking about time, or constantly updating probabilities, well then you can’t escape Bayes: https://www.bayesrulesbook.com/
@Nevermnd thank you for your suggestion , i would surely try it, here’s a look about what i mean data have no seasonality
What courses on LSTMs have you attended (or have experience in their use)?
Few, Some videos on youtube and some articles just that
I recommend you attend Course 5 of the Deep Learning Specialization.
Thank you , i would surely do
@ABOU_ELKHIR_MOHAMED Models are great, and perhaps there is not direct ‘seasonality’, but there definitely appears to be some sort of pattern. But what on Earth transpired Jan First 2023 ? I mean if you just fed this all to a model, I think it would get it wrong. So maybe our ‘wet brains’ are still worth something. I mean only you know what that change was-- But I’d also constrain my dataset from that point forward.
I mean you must know, but any model isn’t going to figure out ‘what happened’.
It simply does not know.
If there is a repeating pattern, an LSTM ought to be able to model it.
@TMosh I am just starting this (final) course presently-- You think even with the past discrepancy ?
I agree, I tried it with that subset of the data and it gave some prediction that arent that far , predict in the 80’s n but still out of usual range , Till now Xgboost gives the best prediction’s but it sould be more accuate
Not sure what you’re referring to.
@TMosh I mean if you look at the graph he presented, up to that date, there is some regularity, but it is a little all over the place. But then it just starts hammering away (with a great deal more variability, yes)-- Though at least it is regular in what is occurring. And @ABOU_ELKHIR_MOHAMED so I come to this space from the high speed printing world at Ricoh, so I have experience working with industrial class machines. So something must have changed then, either in the machine or measurement. Thus, just from a practical point of view, also make sure you are still measuring the relevant variables. I mean at the end of the day, it is not all about models if you are not fitting the process somehow.
Some real-life systems don’t model very well. It’s usually due to situations that are not predictable, or important characteristics that aren’t in the dataset.
Yes , i belive since 2023 they wanted to speed up the production of some elements, which generate more compressed air that is used ro spin a turbine to generate electricity , the data was given to me by the company itself and the mesurements are so accurate, i dont find any problem in the variables to target mapping , all models does so well , but to forecast , that’s a real issues am having
@ABOU_ELKHIR_MOHAMED ok, hopefully someone smarter than me can help you. I only state what seems obvious.
1 Like