Since 2014, HKO has been collaborating with researchers of the Hong Kong University of Science and Technology (HKUST) on the development of new techniques for predicting the evolution of radar images and rainfall using machine learning with recurrent neural network. To date, three algorithms based on convolutional long-short-term-memory (ConvLSTM), convolutional gated recurrent units (ConvGRU), and trajectory gated recurrent units (TrajGRU) have been developed. They are elaborated in the two peer-reviewed papers presented on the international conferences on neural information processing systems (NIPS) in 2015 and 2017. Verification suggests that these novel algorithms outperform the operational variational optical flow algorithm in the tested sample. In time, they may form one of the most important elements of quantitative precipitation forecasts (QPF).
Shi, X., Gao, Z., Lausen, L., Wang, H., Yeung, D. Y., Wong, W. K., & Woo, W. C. (2017). Deep Learning for Precipitation Nowcasting: A Benchmark and A New Model. In Advances in Neural Information Processing Systems (pp. 5622-5632).
Xingjian, S. H. I., Chen, Z., Wang, H., Yeung, D. Y., Wong, W. K., & Woo, W. C. (2015). Convolutional LSTM network: A machine learning approach for precipitation nowcasting. In Advances in neural information processing systems (pp. 802-810).