Improving on the precipitation forecasting of NWP based on convolutional neural network

    Precipitation forecasting is one of the key points and difficulties in weather forecasting. Numerical forecasting is the core means of precipitation forecasting, but its output results are often affected by model uncertainty, uncertainty of initial conditions, errors in parameterization schemes, spatial resolution, topography and geomorphology, and other complex factors, which lead to systematic bias of numerically forecasted precipitation, difficulty in accurately predicting the magnitude of heavy rainfall or sudden heavy precipitation, etc., and therefore, post-processing optimisation of numerical precipitation forecasting is needed. In recent years, studies have shown that compared with the traditional statistical post-processing technology, the artificial intelligence-based post-processing technology in medium-term numerical prediction model has own advantage in that it is a data-driven method. This technology can implicitly extract the spatio-temporal variations of nonlinear and multi-scale physical relationships from multi-source data, thus significantly enhancing the level of medium- and short-term weather forecasts.
    In this paper, a convolutional dendritic neural network (CDNN) model incorporating dendritic (DD) network is constructed on the basis of CNN model, in order to deal with the sample imbalance problem of deep learning precipitation forecast, in which the features are labelled with K-means clustering labels and tag labels.
    A 24-h precipitation forecasting study was carried out using the European Centre for Medium-Range Weather Forecasts Integrated Forecasting System (ECMWF-IFS) forecast product and the  CMA（China Meteorological Administration）Multi-source Precipitation Analysis System（CMPAS）hour-by-hour precipitation analysis product of China. 
    The results show that the CDNN model reduces the average RMSE of each class of precipitation, and has better overall performance in precipitation TS scores, bias, misses, and false alarms, which are closer to the observation. The K-means clustering labels constrain the samples, which allows the CDNN model after K-means clustering labels（K-CDNN） to learn the class information of precipitation, and significantly reduces the RMSE of precipitation forecast, significantly improves the medium and heavy precipitation forecasts. The labelled labels further constrain the samples of the model training, so that the labelled-labels（LK-CDNN） model based on K-means clustering labels takes more into account the small-sample events, and has better prediction ability for heavy precipitation.