Application of Seq2Seq models for predicting the development of thunderstorm activity to enhance the pilot’s situational awareness in flight

Abstract

The paper presents the results of application of Seq2seq models based on neural networks for nowcasting-forecasting with a lead time of up to 2 hours – of thunderstorm activity in order to increase situational awareness of aircraft crews. Various recurrent and convolutional recurrent models were created and trained on the basis of radar meteorological observations of thunderstorm cells. The results showed that convolutional recurrent neural networks (ConvRNN, ConvLSTM, ConvGRU) outperform classical recurrent models and improve the thunderstorm forecast by 25–30% in terms of RMSE (root mean square error) metric compared to the baseline model, which always selects the most recent radar image available at the time of prediction. Nevertheless, despite the fact that the convolution recurrence models can accurately represent the general trend of thunderstorm cloud shape changes, the accuracy of predicting the intensity of thunderstorm cells is usually overestimated. Application of the proposed thunderstorm activity forecasting technology can enhance the situational awareness of the flight crew improving the projection of the current situation into the near future and optimizing the decision-making process for thunderstorm avoidance by providing crew members with predictive information about thunderstorm development on the navigation display screen. Future research is expected to further optimize the model architecture and integrate the predictive technology into flight crew decision support systems.

Keywords

• seq2seq
• prediction
• recurrent neural networks
• convolutional recurrent neural networks
• situational awareness
• thunderstorm avoidance