Overfitting CNN LSTM (Time Series Prediction)

Hi all! I’m doing Time Series Prediction with the CNN-LSTM model, but I got overfitting condition. Here is my model code:

class LSTM(nn.Module):

    def __init__(self, num_classes, input_size, hidden_size, num_layers, seq_length):
        super(LSTM, self).__init__()
        self.num_classes = num_classes
        self.num_layers = num_layers
        self.input_size = input_size
        self.hidden_size = hidden_size
        self.seq_length = seq_length
        self.conv = nn.Conv1d(in_channels=num_classes, out_channels=num_classes, 
                              kernel_size=(3,), padding=2, 
        self.maxpool = nn.MaxPool1d(kernel_size=2, stride=1)
        self.lstm = nn.LSTM(input_size=input_size, hidden_size=hidden_size,
                            num_layers=num_layers, batch_first=True, dropout=0.25)
        self.batchnorm = nn.BatchNorm1d(hidden_size)
        self.fc = nn.Linear(hidden_size, out_features=num_classes)
        self.sig = nn.Sigmoid()

    def forward(self, x):
        x = x.permute(0,2,1)
        conv_out = self.conv(x)
        maxpool_out = self.maxpool(conv_out)
        maxpool_out = maxpool_out.permute(0,2,1)
        h_0 = torch.zeros(self.num_layers, x.size(0), self.hidden_size, requires_grad=True)
        c_0 = torch.zeros(self.num_layers, x.size(0), self.hidden_size, requires_grad=True)

        output_out, (h_out, c_out) = self.lstm(maxpool_out, (h_0, c_0))

        out = output_out[:, -1, :]
        batchnorm_output = self.batchnorm(out)
        linear_output = self.fc(batchnorm_output)
        sigmoid_output = self.sig(linear_output)

        return sigmoid_output

And here is my train code:

seed = 23

num_epochs = 1000
learning_rate = 0.1
momentum = 0.9

input_size = 11
hidden_size = 32
num_layers = 2

num_classes = 11

model = LSTM(num_classes, input_size, hidden_size, num_layers, seq)

loss_func = nn.L1Loss()
print("-" * 70)
validator = True
optim_input = ""
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)

scheduler = ReduceLROnPlateau(optimizer, mode='min', patience=75, verbose=True)

loss_list = []
epochs = []
time_per_epoch = []
# Pelatihan model
print("-" * 70)
print("Begin training with Convolutional LSTM layer (with Batch Normalization) and Adam optimizer\n")
for epoch in range(num_epochs):
    start = time.time()
    outputs = model(train_x)
    loss = loss_func(outputs, train_y)
    stop = time.time()
    epoch_time = stop - start
    print("Epoch: %d, MAE: %1.5f Time: %1.3f s" % (epoch, loss.item(), epoch_time))

print("-" * 70)

print("-" * 70)
print("Training done in %1.5f seconds with last training MAE: %1.5f" % (sum(time_per_epoch), loss_list[-1]))
plt.figure(figsize=(10, 4))
plot_function(time_step=epochs, series=loss_list,
              title='Training Loss Convolutional LSTM Layer with Batch Normalization\nOptimizer: Adam',
              xlabel='Time', ylabel='Loss', label=["loss graph"])

For further information, my train set shape is [764,30,11] that consist of [batch_size,sequence_length,input_size] and my test set shape is [191,30,11]. My dataset is a normalized dataset with MinMaxScaler from scikit-learn. In training, I got an MAE of 0.01567, while in testing, I got an MAE of 0.16626, it’s a very huge difference between the normalized train and test.
Here is the image of the test set:

And here is the image of the prediction of the test set:

What am I supposed to do to reduce this overfitting condition?
Thank you in advance!