Forecasting/prediction multi step ahead with individual events (NARNET Neural Network)

I have around 300 different time series data sets of xyz acceleration during parabolic flights.

I want predict the next 5 second or so using NARnet. But I cannot figure out how to use that data as the examples use one continious data set and not multiple singular events.

Any help would be much appreciated.

Answer 1

[Aditya]

When dealing with multiple time series datasets, such as your 300 different sets of XYZ acceleration during parabolic flights, and you want to predict future values using a Neural Autoregressive Network (NARnet), you'll need to consider how to structure your data for training. Here's how you can approach this problem:

1. Preprocess the data
2. Combine the data : After preprocessing, you can combine the segmented sequences from all your datasets into one large dataset. This combined dataset will have many sequences that the network can learn from. Make sure to shuffle the sequences to prevent the network from learning the order of the flights instead of the underlying patterns.
3. Create and Train the NARnet:

% Assuming 'combinedData' is a cell array where each cell contains a
% preprocessed time series segment (e.g., 5-second long sequences)

% Convert data to a format that NARnet can use
[inputSeries, targetSeries] = preparets(narnet, combinedData, combinedData);

% Create a NARnet
narnet = narnet(inputDelays, hiddenSizes, 'open', trainFcn);

% Train the network
[narnet, tr] = train(narnet, inputSeries, targetSeries);

% 'tr' contains training record, such as the performance and states
% 'lastFewPoints' is a matrix containing the last few points of a time series
% before the 5-second period you want to predict.

% Prepare the data for prediction
[xs, xi, ai, ts] = preparets(narnet, {}, {}, lastFewPoints);

% Predict the next 5 seconds (assuming your time step is 1 second)
predictedValues = narnet(xs, xi, ai);

Remember that the exact implementation details will depend on your specific data and requirements, but this outline should give you a starting point for how to approach training a NARnet with multiple time series datasets.