Deep neural network training for time series forecasting

Question

I'm starting work on predicting Internet traffic (time series prediction) using artificial neural networks, but I have little experience in this matter.

• Does anyone know which method is best for this? (what kind of neural network for time series forecasting)

• Deep Learning with Uncontrolled Learning - Is a Good Idea for a Time Learning Series?

Answer 1

You can do time series prediction with neural networks, but it can become quite complicated.

1) The obvious choice is a recurrent neural network (RNN). However, they are really difficult to train, and I would not recommend RNN if you are using neural networks for the first time. Some interesting work has been done recently to facilitate RNN training (for example, optimization without the Hessian), but then again, this is probably not for beginners ;-) In addition, you can try a scheme in which you use a standard neural network ( i.e. not RNN) and try to predict the next data frame from the previous? That might work.

2) This question is too general, there is no definitive correct answer. Yes, you can use uncontrolled function recognition as part of your decision (for example, pre-train your model), but if your ultimate goal is to predict the time series, you will also need to conduct some controlled training.

Good luck

Answer 2

You can try these documents.

https://arxiv.org/pdf/1803.03916.pdf https://arxiv.org/abs/1905.01697

Common CNNs are state-of-the-art and you should probably try them out.