Learning low-dimensional representations of ensemble forecast fields using autoencoder-based methods