How to use pandas.get_dummies on prediction time

I have to convert categorical value columns to run a deep learning model. In training time I am using pd.get_dummies(data=data, dummy_na=True) , and it is doing completely fine.

But In real-time, I have some new categories in each column as expected. So I have to put it ` column_name_nan' for all of that categories. as get_dummies does not do that, I am written it on my own. I have saved the training time columns in a file and used them in prediction time. But it takes a lot of time for big data like 1M. Here is my code.

def populate_data_frame_in_prediction_time(data, columns):
    unknown_col = "nan"
    columns_set = set(columns)
    result_data_frame = pd.DataFrame(0, index=np.arange(len(data)), columns=columns)

    for prefix in data.columns: # O(m)
        unknown_column_name = str(prefix) + "_" + str(unknown_col)
        for index, row in data.iterrows(): #O(n)
            value = row[prefix]
            result_column_name = str(prefix) + "_" + str(value)
            if result_column_name not in columns_set: # O(1)
                result_column_name = unknown_column_name

            result_data_frame[result_column_name][index] = 1

    result_data_frame = result_data_frame.astype('uint8')
    return result_data_frame

do you guys have a better solution for this?

lets say

    training_data = pd.DataFrame({'City': ['Delhi', None, 'Hyderabad', 'Delhi', 'Hyderabad'],
                         'Country': ['D', 'H', 'D', 'H', None]})
    real_time_data = pd.DataFrame({'City': ['Delhi', None, 'Dhaka'],  # })
                                   'Country': [None, "H", "B"]})
    #encoded_columns_in_training_time_to_use_it_on_prediction_time
{'City': ['City_Delhi', 'City_Hyderabad', 'City_nan'], 'Country': ['Country_D', 'Country_H', 'Country_nan']}