How to forecast with a CNN 2784 setps in the future after the last value in the time series

Dear all, I trained a CNN that works well and validated it with the predict function. Then, I saved the trained model in the file v_1.mat, in the variable "trainedModel". Now, using the trained model, I want to forecast the next 2784 values found imediatelly after the last value in the time series that I used initially for training and validation. The actual data for comparison are in the file 'february_24.xlsx'. In a separate script, I use the folowing code but it says "Incorrect number or types of inputs or outputs for function forecast."

clc

load('v_1.mat');

past_data=xlsread('training_testing_data.xlsx','35233:38016');
yf=forecast(trainedModel,transpose(past_data),2784);
data_february=xlsread('february_24.xlsx','1:2784');
mape(transpose(data_february),yf)
figure(1)
plot(data_february, 'DisplayName', 'Actual');
hold on;
plot(yf, 'DisplayName', 'Predicted');
legend('show');
title('Forecast february 2024');

What am I doing wrong? Is the "forecast" function not compatible to a neural network model ? Is there any alternative to this code ?

Thank you all very much in advance!

Answer 1

[Vinayak]

Hi Adrian,

The documentation for forecast says it accepts a linear/non-linear time-series based system model as input (here). I don’t have your model, but most CNNs are not in such format. I would suggest you use the predict function in a loop refeeding it the latest data for the next ‘n’ values.

The sample code might look something like:

numForecasts = 2784; 
yf = zeros(1, numForecasts);   
inputData = transpose(past_data); 
for i = 1:numForecasts 
    % Predict the next value 
    nextValue = predict(trainedModel, inputData(:, end)); 
    % Append the next value to the forecast array 
    yf(i) = nextValue; 
    % Update the input data by appending the predicted value 
    inputData = [inputData, nextValue]; 
end 

You also might consider creating a custom forecast function which can perform similar calculation under the hood.

I hope this helps!