We propose automatic speech recognition (ASR) models inspired by echo state network (ESN) [1], in which a subset of recurrent neural networks (RNN) layers in the models are randomly initialized and untrained. Our study focuses on RNN-T and Conformer models, and we show that model quality does not drop even when the decoder is fully randomized. Furthermore, such models can be trained more efficiently as the decoders do not require to be updated. By contrast, randomizing encoders hurts model quality, indicating that optimizing encoders and learn proper representations for acoustic inputs are more vital for speech recognition. Overall, we challenge the common practice of training ASR models for all components, and demonstrate that ESN-based models can perform equally well but enable more efficient training and storage than fully-trainable counterparts.
Nowcasting methods based on deep learning typically rely solely on radar data. However, effectively leveraging multi‐source data with diverse spatio‐temporal resolutions remains a significant challenge in the field. To address this challenge, we propose and validate a novel deep learning model for nowcasting, termed Nowcastformer. This model utilizes radar data and upper‐air atmospheric variables, and has been pretrained on satellite data from non‐target regions. Quantitative statistical assessments demonstrate that both the integration of multi‐source data and the implementation of pre‐training strategies enhance the model's performance. Additionally, we conduct a comprehensive analysis of predictor importance, revealing a trend where atmospheric variables become increasingly important as the forecast horizon increases. To illustrate the model's interpretability, we employ the integrated gradients method, which highlights critical areas in representative cases and provides insights into the model's decision‐making process.
Precipitation nowcasting has long been a challenging problem in meteorology. While recent studies have introduced deep neural networks into this area and achieved promising results, these models still struggle with the rapid evolution of rainfall and extremely imbalanced data distribution, resulting in poor forecasting performance for convective scenarios. In this article, we evaluate the amount of information in different precipitation nowcasting tasks of varying lengths using mutual information. We propose two strategies: the mutual information-based reweighting strategy (MIR) and a mutual information-based training strategy (time superimposing strategy (TSS)). MIR reinforces neural network models to improve the forecasting accuracy for convective scenarios while maintaining prediction performance for rainless scenarios and overall nowcasting image quality. The TSS strategy enhances the model’s forecasting performance by adopting a curriculum learning-like method. Although the proposed strategies are simple, the experimental results show that they are effective and can be applied to various state-of-the-art models.
Machine learning (ML) precipitation forecasting models typically employ a mean squared error (MSE) loss function for optimization. However, due to the well‐known “double penalty” effect, MSE‐based losses often lead to overly smoothed prediction fields and a systematic underestimation of the frequency of heavy rainfall. To address this limitation, we propose a novel probability‐matching (PM) based loss for ML precipitation nowcasting and short‐term forecasting, comparing its performance with other classical losses. Comprehensive skill evaluation demonstrates that PM‐based loss offers relatively more balanced and consistent performance across metrics, particularly its lower forecast bias from light to heavy rainfall intensities. Spectral power analysis further indicates that PM‐based loss better preserves small‐scale precipitation variability throughout the forecast period. Additionally, it results in a forecast frequency distribution of precipitation that more closely aligns with the observed distribution. These findings indicate consistent improvements in predictive skill and reliability of ML precipitation forecasts when trained with the PM‐based loss.
In this paper, we explore the prevalent issue of underestimation of extreme precipitation values in deep learning models utilized for precipitation forecasting. We emphasize that this challenge arises from the double penalty phenomenon, which is exacerbated by the joint effect of the commonly adopted mean squared error (MSE) loss function and the intrinsic uncertainty of forecasting tasks. Drawing inspiration from probability-matching ensemble forecasting, we introduce Sort Loss, a straightforward yet highly effective deep learning loss function. By leveraging the ordinal relationships within meteorological data, Sort Loss circumvents the positional information-related double penalty problem. Experimental results from precipitation nowcasting and short-term forecasting tasks demonstrate that Sort Loss effectively diminishes the distributional discrepancies between model forecasts and actual observations. Consequently, it significantly enhances forecasting performance in heavy rainfall scenarios, while simultaneously maintaining stability across other weather conditions. This study offers a novel perspective on optimizing deep learning models for weather forecasting and showcases the potential of applying Sort Loss to improve the accuracy of extreme weather predictions.