Multi-step prediction code construction

• Constructing the multi-step prediction model is my target.
• 17 channels of input and 17 channels of output exist in my dataset.
• Specific period (ex. 7 days) of input and output data is utilized for training the model.
• My Q. Is the code is well constructed for multi-step prediction model for same period of input and output?
• If the length of input and output data is different, the Code does not works. How can I modify the code? I think if I change Y_train, cell into double array, num of output layer cell should be increased in (num_channels * output target time steps). But this method is not proper.

x_train = rand(9665,17); % data at T for 17 channels (For Training)
y_test = rand(9665,17); % data at T+1 for 17 channels (For Training)

x_test = rand(2417,17); % data at T for 17 channels (For Testing)
y_test = rand(2417,17); % data at T+1 for 17 channels (For Testing)

%% Multi-step ahead prediction with adjustable input/output windows
% input_steps and output steps are same now (assumption)
input_steps = 1000; % 1주 (24시간, 시간 간격 6개)
output_steps = 1000;       % 1시간 (6개 간격)


num_samples_train = numtrain - input_steps - output_steps + 1;
num_samples_test  = size(x_test,1) - input_steps - output_steps + 1;

X_train = cell(num_samples_train,1);
Y_train = cell(num_samples_train,1);

X_test  = cell(num_samples_test,1);
Y_test  = cell(num_samples_test,1);

%% 셀 배열에 각 시퀀스를 저장
for i = 1:num_samples_train
    % X: (input_steps × 모든 채널)
    X_train{i} = x_train(i : i+input_steps-1, :);
    % Y: (output_steps × Tension 채널 수)
    Y_train{i} = Tension(i+input_steps : i+input_steps+output_steps-1, :);
end

for i = 1:num_samples_test
    X_test{i} = x_test(i : i+input_steps-1, :);
    Y_test{i} = Tension(i+input_steps : i+input_steps+output_steps-1, :);
end

f = figure;
hold on
plot(1:1000,X_train{44}(:,1),'--')
plot(1001:2000,Y_train{44}(:,1),'-')

%% 각 셀의 크기 확인 (예시)
disp(size(X_train));  % [8514, 1] 형태
disp(size(X_train{1}));  % [1008, 36] 형태 (예시)
disp(size(Y_train{1}));  % [144, 17]  형태 (예시)
%% model architecture
net = dlnetwork;

tempNet = [
    sequenceInputLayer(17,"Name","sequenceinput")
    batchNormalizationLayer("Name","batchnorm")
    convolution1dLayer(3,32,"Name","conv1d_4","Padding","same")
    reluLayer("Name","relu_5")];
net = addLayers(net,tempNet);

tempNet = identityLayer("Name","identity");
net = addLayers(net,tempNet);

tempNet = [
    convolution1dLayer(4,32,"Name","conv1d","Padding","same")
    maxPooling1dLayer(3,"Name","maxpool1d","Padding","same")
    reluLayer("Name","relu")];
net = addLayers(net,tempNet);

tempNet = [
    additionLayer(2,"Name","addition")
    reluLayer("Name","relu_6")
    identityLayer("Name","identity_1")];
net = addLayers(net,tempNet);

tempNet = identityLayer("Name","identity_2");
net = addLayers(net,tempNet);

tempNet = [
    convolution1dLayer(32,32,"Name","conv1d_2","Padding","same")
    maxPooling1dLayer(3,"Name","maxpool1d_2","Padding","same")
    reluLayer("Name","relu_2")];
net = addLayers(net,tempNet);

tempNet = [
    additionLayer(2,"Name","addition_2")
    reluLayer("Name","relu_7")
    identityLayer("Name","identity_3")];
net = addLayers(net,tempNet);

tempNet = [
    convolution1dLayer(4,32,"Name","conv1d_1","Padding","same")
    maxPooling1dLayer(3,"Name","maxpool1d_1","Padding","same")
    reluLayer("Name","relu_1")];
net = addLayers(net,tempNet);

tempNet = [
    additionLayer(2,"Name","addition_1")
    reluLayer("Name","relu_8")];
net = addLayers(net,tempNet);

tempNet = [
    convolution1dLayer(32,32,"Name","conv1d_3","Padding","same")
    maxPooling1dLayer(3,"Name","maxpool1d_3","Padding","same")
    reluLayer("Name","relu_3")];
net = addLayers(net,tempNet);

tempNet = [
    additionLayer(2,"Name","addition_3")
    reluLayer("Name","relu_9")];
net = addLayers(net,tempNet);

tempNet = [
    additionLayer(2,"Name","addition_4")
    % functionLayer(@(X) cbt2cb(X), Formattable=true, Acceleratable=true)    
    reluLayer("Name","relu_10")
    fullyConnectedLayer(128,"Name","fc")
    reluLayer("Name","relu_4")
    fullyConnectedLayer(17,"Name","fc_1")];
net = addLayers(net,tempNet);

% 헬퍼 변수 정리
clear tempNet;

net = connectLayers(net,"relu_5","identity");
net = connectLayers(net,"relu_5","identity_2");
net = connectLayers(net,"identity","conv1d");
net = connectLayers(net,"identity","addition/in1");
net = connectLayers(net,"relu","addition/in2");
net = connectLayers(net,"identity_2","conv1d_2");
net = connectLayers(net,"identity_2","addition_2/in1");
net = connectLayers(net,"relu_2","addition_2/in2");
net = connectLayers(net,"identity_1","conv1d_1");
net = connectLayers(net,"identity_1","addition_1/in1");
net = connectLayers(net,"relu_1","addition_1/in2");
net = connectLayers(net,"identity_3","conv1d_3");
net = connectLayers(net,"identity_3","addition_3/in1");
net = connectLayers(net,"relu_3","addition_3/in2");
net = connectLayers(net,"relu_8","addition_4/in1");
net = connectLayers(net,"relu_9","addition_4/in2");
net = initialize(net);

analyzeNetwork(net);
%%
options = trainingOptions('adam', ...
    'MaxEpochs',10, ...
    'InitialLearnRate',0.005, ...
    'GradientThresholdMethod','global-l2norm',...
    'L2Regularization',1e-5,...
    'MiniBatchSize',50,...
    'GradientThreshold',1,...
    'Shuffle','every-epoch',...
    'LearnRateSchedule','piecewise',...
    'LearnRateDropPeriod',100,... 
    'VerboseFrequency',50,...
    'LearnRateDropFactor',0.9,...
    'ExecutionEnvironment','auto',...
    'Verbose',true, ...
    'Plots','none');

% 모델 학습
net1 = trainnet(X_train,Y_train, net, 'mse', options);

%% Model Testing
Y_pred = minibatchpredict(net, X_test);

%%
f=figure;
hold on
plot(Y_pred(:,1,1))
plot(Y_pred(1,:,1))

Answer 1

[TED MOSBY]

Hi Jingoo,

Your final network layer is fullyConnectedLayer(17):

• That means each mini‐batch sample ends up being mapped to a 17‐dimensional output vector (i.e. one time step of 17 channels).
• However, your Y_train{i} and Y_test{i} cells each contain output_steps time steps × 17 channels.
• So, if output_steps > 1, the network output dimension does not match the dimension of Y_train{i} any more.

Flatten the time dimension at the very end and use a single fullyConnectedLayer that outputs all (output_steps × number_of_channels) at once. After the network outputs that vector, you reshape it back to [output_steps, number_of_channels].

Make sure that whatever training function (trainnet, minibatchpredict, etc.) you are using can handle your new shape of Y_train{i} and does the MSE properly.

Hope this helps!