Keras' model.predict() produces very different accuracy when calling on multiple batches at once VS calling on individual batches one by one

This is the model I've used:

model = Sequential()
model.add(LSTM(units=200, input_shape=(15, 17), return_sequences=True)
model.add(Dropout(drop_out))
model.add(BatchNormalization())
model.add(LSTM(units=unit_per_layer))     
model.add(Dropout(drop_out))
model.add(BatchNormalization())
model.add(Dense(units=unit_per_layer, activation='tanh'))
model.add(Dropout(drop_out))
model.add(BatchNormalization())
model.add(Dense(units=1, activation='sigmoid'))

model.compile(optimizer=opti_func, loss='binary_crossentropy', metrics=['binary_accuracy'])

Now, after model.fit(), when I call model.predict(X_dataset_multiple_batch) I get some good predictions. However, if I devide X_dataset_multiple_batch into a series of individual batches (let's call each of them X_dataset_single_batch) and call model.predict(X_dataset_single_batch) on them one by one (i.e. call model.predict(X_dataset_single_batch) multiple times), the predictions become much worse than the former one.

Additional Note Honestly I only need the last prediction, but because of Keras internal design stuff, I cannot have only one single prediction. It has to be in batches. So I have to do my predictions on batches and then extract the last prediction. This is fine, but the question now becomes what is the optimum number of batches that I have to give model.predict()?

Update 1

This is how I pick individual batches from my test.csv:

df = pd.read_csv("/content/drive/MyDrive/Colab Notebooks/Data/Test_file.csv")

X_ = df.loc[:, 'b':'dm'].to_numpy()
Y_ = df.loc[:, 'dn'].to_numpy()

predictions = []

for i in range (0, 2 + X_.shape[0] - (n_batch + time_step)):
    X = X_[i:i+n_batch+time_step]
    Y = Y_[i:i+n_batch+time_step]

    scaler = MinMaxScaler()
    X = scaler.fit_transform(X)

    X_one_batch_of_data = np.zeros((n_batch, time_step, X.shape[1]))
    
    for j in range(0, X_one_batch_of_data.shape[0]):
        X_one_batch_of_data[j] = X[j:j+time_step, :]

    predictions.append(model.predict(X_one_batch_of_data, batch_size=n_batch)[-1, 0])

from Keras' model.predict() produces very different accuracy when calling on multiple batches at once VS calling on individual batches one by one