Study On Spatiotemporal Feature Learning Method Of Radar Echo Sequence Prediction For Short Time Approaching Precipitation

Precipitation forecasting plays a vital role in daily life.From agricultural irrigation to people’s production and service,precipitation forecasting has its shadow.In recent years,with the rise of deep learning,precipitation forecasting methods have gradually transferred from the previous numerical weather forecasting methods to deep learning methods.Spatiotemporal series prediction is one of the important research directions for short-term near precipitation prediction,which is more accurate than numerical weather prediction.When conducting spatiotemporal series prediction tasks,spatiotemporal units and the structural network built by them affect the accuracy of precipitation near prediction results.In the prediction of radar echoes,the space-time unit usually has fuzzy prediction results due to the short-term mutation phenomenon of the sequence and its own catastrophic forgetting,and the space-time prediction network will lose a large amount of important information in the process of information transmission,resulting in deformation and ambiguity of the prediction results.Therefore,this paper has improved the space-time unit and space-time network in the current space-time series prediction process.The ability of spatio-temporal unit and spatio-temporal network to capture spatio-temporal information is optimized to improve the accuracy of precipitation prediction.The main work of this paper is as follows:(1)This paper proposes a universal enhanced gating mechanism that improves the ST-LSTM unit commonly used in spatio-temporal prediction tasks by using gating mechanisms.Firstly,the traditional gating mechanism is extended to the radar echo prediction sequence image,which effectively delays the saturation phenomenon in the ST-LSTM unit and improves the prediction accuracy.Secondly,to address the unique short-term information mutation problem in the spatiotemporal domain,we further propose an enhanced gating mechanism,which enhances the STLSTM unit’s ability to capture short-term mutation information,strengthens the extraction of important information from the radar echo sequence,and names the spatio-temporal unit with enhanced gating mechanism as EnTG-ST-LSTM unit.Finally,we use this unit to build a traditional prediction structure to achieve multi-step prediction of radar echo sequences and improve the prediction effect of short-term precipitation.(2)Catastrophic forgetting often occurs in the forget gate of the traditional spatiotemporal prediction unit ST-LSTM,leading to the loss of long-term memory information and making it difficult to predict the spatiotemporal sequence results.To solve this problem,we propose the SA-ST-LSTM unit based on the ST-LSTM unit,which replaces the forget gate mechanism in the unit module to avoid catastrophic forgetting and improve the network’s prediction ability.Secondly,in the prediction process,we usually use the encoder-decoder structure for multi-step prediction.We introduce the self-attention mechanism and use a new encoder--attention-decoder mechanism for prediction,which not only effectively preserves long-term memory information during transmission but also strengthens the extraction of important information in the image.In the stacking process,we use a new way of stacking CNN and RNN to encode and decode information in two layers and obtain better prediction results.In this paper,MovingMNIST simulation dataset,KTH pedestrian dataset and CKIM radar echo dataset were used to verify the spatiotemporal series prediction effect of the improved algorithm.In the experiment,the two algorithms designed in this paper obtained good prediction effect on the three datasets.By comparing the PredRNN network constructed by ST-LSTM unit as the baseline network,the HSS index of the first algorithm is 4%higher than that of the baseline network when the threshold is 30,and the HSS index of the second algorithm is 7.9%higher than that of the baseline network when the threshold is 30.This shows that the two algorithms can effectively improve the radar echo sequence prediction results,and judge the intensity and region of precipitation prediction more accurately.