Getting output values from Neural net time series for energy price prediction

I am currently using Matlab Neural net time series to predict future values of energy prices for school work.

I have simply x and y values.

X values are actually dates, which I have transformed into numeric values by simply assigning 1 for the first date data. So they go as 1,2,3,... from 1 to 48.

Dependent y values are numeric energy price values such as 7,10,15, etc.

I am using Nonlinear input-output network for my analysis.

After I export model as Network Function to MATLAB Coder, I believe I should be using

[y1,xf1] = myNeuralNetworkFunction(x1,xi1)

to get outputs from trained network.

I have 48 data, and I want to use trained network to predict 49,50,...68 data until 68th.

How am I supposed to continue?

When I input [49] for the x1 value to get prediction of 49th value, I realized that input of xi1 (delay input) affects the output significantly. What should I write to xi1 to get a good result?

Answer 1

[Ayush Aniket]

Since you are using a Nonlinear Input-Output Time Series Network, the xi1 input is crucial for accurate forecasting beyond your training data. Here’s how to proceed:

1. Understand xi1 (Delayed Input States) - The network uses past input values to predict future outputs. xi1 holds the delay history, typically the last d values of your input series, where d is the number of delays used during training.

If your network was trained with, say, 2 delays, then to predict the 49th value, you need:

x1 = {49};             % Current input
xi1 = {[47 48]};       % Delayed inputs (from training data)

These delayed values help the network maintain context.

2. Predict Multiple Future Steps - To predict values from 49 to 68, you need to simulate forward iteratively, updating xi1 with each new prediction:

% Initialize with last known inputs
xi = {[46 47 48]};  % Example for 3 delays

for t = 49:68
    x1 = {t};  % Current time step
    [y1, xf1] = myNeuralNetworkFunction(x1, xi);
    
    % Store prediction
    predictions(t) = y1{1};
    
    % Update xi for next step
    xi = [xi{1}(2:end), x1];  % Shift and append new input
    xi = {xi};                % Wrap in cell
end

This loop feeds each new prediction back into the delay line, simulating the network forward.

Points to note:

• The quality of xi1 directly affects prediction accuracy. You should use actual historical values for the first prediction, then update with predicted values.
• If your network was trained in open-loop mode, consider converting it to closed-loop using closeloop(net) before exporting. This allows it to use its own predictions as future inputs.