I am doing time series forecasting using neural networks with multiple input and single output. My data have both trend and cyclical components, I use two neurons in hidden layer and one in output layer. The forecasting result of each step ahead have obviously delay than it's target, How can I decrease these ANN forecasting delays?
My experience with ANN involves the development of binary classifiers, not time series prediction. However, it's interesting that you only have 2 neurons in the hidden layer. Recall that you can approximate any function with a single hidden layer, provided you have enough neurons in the hidden layer. Maybe you need to increase the number of neurons in the hidden layer (assuming you have enough data to avoid the curse of dimensionality)? Or, maybe perform non-linear feature selection to remove redundant or irrelevant features/inputs which otherwise can make it harder to approach the global minimum during training?
Ussually, when you want to predict time series you need a network with some form of "memory" like recurrent neural networks. they have better results than feed-forward neural netiworks.
regards,
Bogdan
Dear Nafas,
What do you mean by "forecasting result of each step ahead have obviously delay than it's target"? Could you explain this? What are ANN inputs and output?
Bogdan: I am not sure that recurrent NNs give better results than FF NN. Can you give some examples of the papers where this is shown?
Best regards
@ Dudek
My Inputs = time series data (decimal number)
My Outputs = one step ahead forecasting of My Inputs (also decimal number)
Delay = one step ahead outputs(e.g blue point2 in fig) than target(e.g red point1 in fig)
the red line is original time series and the blue line is my forecasting results.
@Oancea
recurrent neural networks may have better results, but I don't want using that, I am looking for another solution for prevention of ANN forecasting delay.
So to improve results you can:
1. Increase lag
2. Add more neurons
3. Work on patterns of time series fragments - see
Article Pattern similarity-based methods for short-term load forecas...
Article Neural Networks for Pattern-Based Short-Term Load Forecastin...
@Nafas, have you successfully resolved the delay problem in forecasting !!!
As I am having same problem with my model
@Sanam
yes, try @Dudek suggestions
1. Increase lag
2. Add more neurons
both suggestions solved my ANN forecasting delay
Thankyou Nafas. I have already attempted for this, I increased the lag terms from 2 to 4 and I added some external variables as input along with the lag terms. The predictions are excellent for the next step ahead but if I try to model for increasing horizon for example for the next 10 steps ahead.... It encounters delay problem in forecasting
I see
my before ANN model did one-step ahead forecasting,
I have the delay challenge in long-term forecasting too, I am studying about patterns of time series fragments for solving that!
If your ANN model is not recurrent plz look @Dudek paper suggestions in this topic
- Badges
- Science topic