Deep Network Designer for time series forecasting

I tried the chickenpox dataset example (https://www.mathworks.com/help/deeplearning/ug/time-series-forecasting-using-deep-network-designer.html) using my own data (single time series first) and went very well.

But I have 4 time series of data and I want to train the network to predict the 4th one. This 4th timeseries depends on the 3 others, so I believe that accuracy will be increased if I use them all. I tried to make a matrix (4 x timesteps) as XTrain and for YTrain (1 x timesteps) I used the 4th row. In deep network designer I put in input layer input size = 4 and in fc I put output=1. An error message informed me that I give 1 input but the network expects 4. After that I put input=1 but I'm not sure what I was doing.

Is there an example with multi input and single output? In the time series forecasting (https://www.mathworks.com/help/deeplearning/ug/time-series-forecasting-using-deep-learning.html?searchHighlight=time%20series%20forecast&s_tid=srchtitle_support_results_1_time%20series%20forecast) the channels are confusing me.

What modification could I make in the chickenpox example to give deep network designer 4 time series and ask to predict the 4th?

Answer 1

[Jaimin]

Hi @Nestoras Papadopoulos

To modify the chickenpox example for a multi-input, single-output time series prediction task, you must configure your input and network architecture to effectively handle multiple input features (the initial three time series) and generate a single output (the fourth time series).

Kindly refer to the following network architecture for understanding.

• Sequence Input Layer: Set InputSize to 3.
• LSTM Layer: Use a suitable number of hidden units (e.g., 50).
• Fully Connected Layer: Set the number of outputs to 1.
• Regression Layer: For predicting continuous values.

Kindly refer to the following code snippet to understand the implementation of the above architecture.

% Example data preparation
XTrain = [TS1; TS2; TS3];
YTrain = TS4;

% Define the network architecture
layers = [
    sequenceInputLayer(3) % 3 input features
    lstmLayer(50, 'OutputMode', 'sequence') % Example LSTM layer
    fullyConnectedLayer(1) % Output size of 1
    regressionLayer
];

% Training options
options = trainingOptions('adam', ...
    'MaxEpochs', 100, ...
    'MiniBatchSize', 20, ...
    'Plots', 'training-progress', ...
    'Verbose', 0);

% Train the network
net = trainNetwork(XTrain, YTrain, layers, options);

% Predict and evaluate
YPred = predict(net, XTest); % Replace XTest with your test data

For more information about “lstmLayer” kindly refer following MathWorks documentation.

https://www.mathworks.com/help/deeplearning/ref/nnet.cnn.layer.lstmlayer.html

I hope this will be helpful.