How to Recursively Predict a Time Series Using Neural Networks

Question

I am currently using neural networks to forecast an electrical demand time series. I am trying to create a forecast for the following day given previous observations at half hourly intervals.

My attempt is to use the previous five day values (48*5=240 half hour intervals) in order to predict predict the next value. Then I wish to use the previous 239 true load values and my predicted load value in order to get another prediction, and so on.

I currently have a model which is shown in the following diagram. However the error is far too high. I am wondering if this is the best way of going about it and if so how to improve on my neural network.

Answer 1

There are different ways to forecast time series with machine learning models. No way is generally better than another, so it is necessary to carry out a trial and error process (a.k.a. time series cross-validation) to determine which is the best method for the specific dataset you are using. In the particular case of neural networks, appropriate feature (variable) selection is crucial for accurate forecasting. So try to use an approach that identifies a set of candidate features based on the data characteristics and then select a subset of them using correlation or instance-based feature selection methods. I recommend you read this excellent thesis so that you have a broader view of the methodologies you can use for this problem.

Answer 2

Well, you are right about the way to make recursive predictions. Maybe there's something wrong with the type of model you've fitted.

Neural networks are very complex models (thus, overfitting machines) and, what's worse, they are black-boxes, so very hard to validate (aparte from just measuring its performance) I would strongly suggest you to try an ARIMA approach unless there is a good reason why you shouldn't