[Issue] How can I apply 'predictAndUpdateState' on a multivariate multistep problem?

Hello,

I have this code below (also attached with data samples) of a multivariate LSTM neural network model.

I am trying to modify it to forecast new steps in the future, but I get this error:

Conversion to double from cell is not possible.
Error in LSTM_multi_motores (line 82)
    [net, Yforecast(:,i)] = predictAndUpdateState(net, XTest(:,i), 'ExecutionEnvironment', 'cpu'); 

I tried to modify this line from

[net, Yforecast(:,i)] = predictAndUpdateState(net, XTest(:,i), 'ExecutionEnvironment', 'cpu'); 

to

[net, Yforecast(:,i)] = predictAndUpdateState(net, XTest{:,i}, 'ExecutionEnvironment', 'cpu'); 

but I got a different error here:

Index in position 2 exceeds array bounds (must not exceed 389).
Error in LSTM_multi_motores (line 82)
    [net, Yforecast(:,i)] = predictAndUpdateState(net, XTest{:,i}, 'ExecutionEnvironment', 'cpu'); 

Could you please tell me how to fix it?

Here's the code:

%% Multivariate Multi-step LSTM NN Model 
% From: Marcelo Olmedo -- MathWorks | Edited by: Mohamed Nedal 
close all; clear; clc 

%% Import data 
[~,~,data_train] = xlsread('ET_1.xlsx'); 
data_mat_train = cell2mat(data_train); 

XTrain = (data_mat_train(:, 4:8))'; 
YTrain = (data_mat_train(:, 3))'; 

XTrain = num2cell(XTrain, 1); 
YTrain = num2cell(YTrain, 1); 

miniBatchSize = 1; % one predictor sequence

%% Define Network Architecture 
numResponses = size(YTrain{1}, 1); 
featureDimension = size(XTrain{1}, 1); 
fprintf('Number of input features: %d \n', featureDimension) 
fprintf('Number of target features: %d \n', numResponses) 

maxepochs = 100; 
miniBatchSize = 1; 
numHiddenUnits = 100; 
% A fully connected layer multiplies the input by a weight matrix and then
% adds a bias vector 
outputSize = 500; 

layers = [ ...
    sequenceInputLayer(featureDimension)
    lstmLayer(numHiddenUnits, 'OutputMode', 'sequence')
    fullyConnectedLayer(outputSize) 
    dropoutLayer(0.1) 
    fullyConnectedLayer(numResponses)
    regressionLayer]; 

options = trainingOptions('adam', ... 
    'MaxEpochs', maxepochs, ...
    'GradientThreshold', 1, ...
    'InitialLearnRate', 0.005, ...
    'LearnRateSchedule', 'piecewise', ...
    'LearnRateDropPeriod', 125, ...
    'LearnRateDropFactor', 0.2, ...
    'Plots', 'training-progress'); 

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

%% Test the Network
[~,~,data_test] = xlsread('ET_2_m_input.xlsx'); 
data_mat_test = cell2mat(data_test); 

XTest = (data_mat_test(:, 4:8))'; 
YTest = (data_mat_test(:, 3))'; 

XTest = num2cell(XTest, 1); 
YTest = num2cell(YTest, 1); 

net = resetState(net); 
YPred = predict(net, XTest); 

rmse = sqrt(mean((cell2mat(YPred)' - cell2mat(YTest)).^2)); 
fprintf('RMSE = %.2f \n', mean(rmse)) 

%% Plot 
y1 = (cell2mat(YTest(1:end, 1:end))'); 
y2 = (cell2mat(YPred(1:end, 1:end)));  % have to transpose as plot plots columns 
figure 
plot(y1) 
hold on 
plot(y2) 
legend('Test', 'Predict') 

%% Forecast the Future 
clc 
net = resetState(net); 
Yforecast = []; 
numTimeStepsTest = numel(XTest) + 500; % forecast new 500 steps in the future 

k = 1; 
for i = 1:numTimeStepsTest 
    [net, Yforecast(:,i)] = predictAndUpdateState(net, XTest(:,i), 'ExecutionEnvironment', 'cpu'); 
    
    if k ~= numel(XTest) 
        [net, Yforecast(:,i)] = predictAndUpdateState(net, XTest(:,i), 'ExecutionEnvironment', 'cpu'); 
    else
        [net, Yforecast(:,i)] = predictAndUpdateState(net, Yforecast(:,i-1), 'ExecutionEnvironment', 'cpu');
    end 
    k = k + 1; 
    
end

fprintf('New data is predicted successfully \n') 

[Xymbu]

Hey did you ever figure out how to address this problem? I am having the ame one

Answer 1

[拓 李]

I suggest you modify it with the help of MATLAB. In addition, how can your time step input be the later time step and output be the previous time step?