Input Size Mismatch on LSTM model development

Good day,

Given that the input to the LSTM has three dimensions: num_features, num_time_steps as 24, and num_training_samples. i intend to predict a 24 hours day-ahead GHI. X_train_3D = 4xs4x1461; X_test_3D = 4x24x365; y_train_reshaped = 24x1x1461.

Below are the codes:

best_rmse = inf;

best_mae = inf;

best_r2 = -inf;;

best_hyperparameters = [];;

% Perform grid search

for optimizer_val = optimizers

for num_lstm_layers_val = num_lstm_layers

for num_hidden_layers_val = num_hidden_layers

for hidden_units_val = hidden_units

for learning_rate_val = learning_rates

for num_epochs_val = num_epochs

for dropout_rate_val = dropout_rates

for batch_size_val = batch_sizes

% Build LSTM model

layers = [ ...

sequenceInputLayer([num_features num_time_steps 1],'Name', 'input')

];

for i = 1:num_lstm_layers_val

layers = [layers

lstmLayer(hidden_units_val,'OutputMode','sequence')

dropoutLayer(dropout_rate_val)

];

end

layers = [layers

fullyConnectedLayer(num_time_steps)

regressionLayer];

% Specify options for training

options = trainingOptions(optimizer_val{1}, ...

'LearnRateSchedule','piecewise', ...

'LearnRateDropPeriod',10, ...

'LearnRateDropFactor',0.9, ...

'InitialLearnRate',learning_rate_val, ...

'MaxEpochs',num_epochs_val, ...

'MiniBatchSize',batch_size_val, ...

'SequenceLength','longest', ...

'Shuffle','never', ...

'Verbose',0, ...

'Plots','training-progress');

% Train LSTM model

net = trainNetwork(X_train_3D, y_train_3D, layers, options);

% Make predictions on test set

y_pred = predict(net, X_test_3D);

% Denormalize predicted GHI values

y_pred_denormalized = y_pred * range(features.GHI) + min(features.GHI);

% Compute evaluation metrics

rmse = sqrt(mean((y_test - y_pred_denormalized).^2));

mae = mean(abs(y_test - y_pred_denormalized));

r2 = 1 - sum((y_test - y_pred_denormalized).^2) / sum((y_test - mean(y_test)).^2);

% Update best hyperparameters if metrics improve

if rmse < best_rmse

best_rmse = rmse;

best_mae = mae;

best_r2 = r2;

best_hyperparameters = [optimizer_val{1}, num_lstm_layers_val, num_hidden_layers_val, hidden_units_val, learning_rate_val, num_epochs_val, dropout_rate_val, batch_size_val];

end

end

end

end

end

end

end

end

end

Error using trainNetwork (line 184)

Invalid network.

Caused by:

Layer 2: Input size mismatch. LSTM layers must have scalar input size, but input size (4×24×1) was received. Try

using a flatten layer before the LSTM layer.

when i first used sequenceInputLayer([num_features num_time_steps],'Name', 'input'), it displayed

Error using sequenceInputLayer (line 131)

Invalid argument at position 1. Expected input to be a scalar, 3 element vector or 4 element vector.'

then I included flattenedLayer before the LSTM layer, it displayed

;Error using trainNetwork (line 184)

Number of elements must not change. Use [] as one of the size inputs to automatically calculate the appropriate size

for that dimension.

Caused by:

Error using reshape

Number of elements must not change. Use [] as one of the size inputs to automatically calculate the appropriate

size for that dimension.

Answer 1

[Cris LaPierre]

The input to the lstmLayer is the number of features (a scaler), not the size of the input.

See this answer: https://www.mathworks.com/matlabcentral/answers/2022292-lstm-layer-input-size

[Babajide Osonuga]

Thanks so much.