Research On Short-term And Impending Precipitation Forecast Based On ConvLSTM Depth Spatiotemporal Network

Precipitation is one of the most common weather phenomena in life.The phenomenon that the water vapor in the atmosphere falls to the ground after condensation is precipitation.Accurate precipitation forecast can not only guide human production and life,but also prevent natural disasters caused by precipitation,so precipitation forecast is of great significance.Due to the uneven spatial distribution and unstable temporal variation of precipitation,precipitation forecast has always been a very challenging problem.Compared with long-term,medium-term and short-term precipitation forecast,short-term and impending precipitation forecast has shorter forecast time limit and higher requirements for forecast accuracy.At present,the most commonly used data in the field of precipitation forecast is meteorological radar echo data.From the radar echo image,we can not only obtain the precipitation in the current time area,but also analyze the movement and change trend of precipitation clouds.Through the radar echo extrapolation technology,we can also predict the future precipitation.Compared with the actual image,the radar echo image extrapolated by the existing radar echo extrapolation technology has the problems of echo loss and inaccurate echo moving track,which leads to the low accuracy and accuracy of precipitation forecast.Because the radar echo data has the correlation of time and space,the extrapolation of radar echo essentially belongs to the problem of spatiotemporal sequence forecast.Convolutional Long Short Term Memory(Conv LSTM)neural network can fully consider the spatiotemporal correlation between data and perform well in spatiotemporal sequence prediction.Therefore,it is often used to solve the problem of short-term and imminent precipitation prediction.Based on the improvement of Conv LSTM deep neural network,this paper adds the Convolutional Block Attention Module(CBAM)to the original network,and combined with the more flexible swats(Switch from Adam to SGD)optimization algorithm,proposes a more effective new rainnet algorithm for short-term and imminent precipitation prediction.A "Encoding-forecasting" network for radar echo extrapolation is constructed by stacking new rainnet modules.The short-term and impending precipitation forecast in the next 0-60 minutes is successfully realized on the radar echo data set of Hong Kong Observatory,and good results are obtained.Experiments show that the network surpasses other methods in both network convergence and precipitation forecast accuracy and correct rate.In addition,the radar echo image generated by the network prediction is also better than other methods in visual effect,retains the image details,and is closer to the real radar echo image than other methods.