WO2023103587A1 - Imminent precipitation forecast method and apparatus - Google Patents
Imminent precipitation forecast method and apparatus Download PDFInfo
- Publication number
- WO2023103587A1 WO2023103587A1 PCT/CN2022/125568 CN2022125568W WO2023103587A1 WO 2023103587 A1 WO2023103587 A1 WO 2023103587A1 CN 2022125568 W CN2022125568 W CN 2022125568W WO 2023103587 A1 WO2023103587 A1 WO 2023103587A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- precipitation
- network
- convlstm
- radar
- short
- Prior art date
Links
- 238000001556 precipitation Methods 0.000 title claims abstract description 59
- 238000000034 method Methods 0.000 title claims abstract description 37
- 238000012549 training Methods 0.000 claims abstract description 15
- 230000007246 mechanism Effects 0.000 claims abstract description 13
- 238000007781 pre-processing Methods 0.000 claims abstract description 4
- 230000015654 memory Effects 0.000 claims description 8
- 230000004913 activation Effects 0.000 claims description 6
- 230000006870 function Effects 0.000 claims description 6
- 230000001186 cumulative effect Effects 0.000 claims description 5
- 238000004590 computer program Methods 0.000 claims description 4
- 238000002592 echocardiography Methods 0.000 claims description 4
- 230000007704 transition Effects 0.000 claims description 4
- 239000003086 colorant Substances 0.000 claims description 3
- 239000011159 matrix material Substances 0.000 claims description 3
- 238000009825 accumulation Methods 0.000 claims 1
- 230000008859 change Effects 0.000 abstract description 4
- 230000000694 effects Effects 0.000 abstract description 4
- 230000000007 visual effect Effects 0.000 abstract 1
- 230000006872 improvement Effects 0.000 description 7
- 230000008569 process Effects 0.000 description 6
- 238000013528 artificial neural network Methods 0.000 description 5
- 238000010586 diagram Methods 0.000 description 3
- 230000005284 excitation Effects 0.000 description 3
- 230000006403 short-term memory Effects 0.000 description 3
- 230000007787 long-term memory Effects 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 230000000306 recurrent effect Effects 0.000 description 2
- 230000004044 response Effects 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 230000004931 aggregating effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000006835 compression Effects 0.000 description 1
- 238000007906 compression Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S13/00—Systems using the reflection or reradiation of radio waves, e.g. radar systems; Analogous systems using reflection or reradiation of waves whose nature or wavelength is irrelevant or unspecified
- G01S13/88—Radar or analogous systems specially adapted for specific applications
- G01S13/95—Radar or analogous systems specially adapted for specific applications for meteorological use
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S7/00—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
- G01S7/02—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00
- G01S7/41—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section
- G01S7/411—Identification of targets based on measurements of radar reflectivity
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S7/00—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
- G01S7/02—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00
- G01S7/41—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section
- G01S7/417—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section involving the use of neural networks
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01W—METEOROLOGY
- G01W1/00—Meteorology
- G01W1/10—Devices for predicting weather conditions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Definitions
- the invention belongs to the technical field of image processing, and in particular relates to a short-imminent precipitation prediction method based on ConvLSTM and SENet attention mechanism through radar echo images.
- Intelligent short-term precipitation forecast generally refers to the precipitation weather forecast within six hours, and the forecast accuracy can reach the kilometer level and minute level. Compared with mid-to-long-term and short-term precipitation forecasts, short-imminent forecasts have higher requirements in terms of demand area, forecast timeliness and other elements. In recent years, changes in marine climate and the frequency and intensity of extreme weather events have been increasing, and the impact of extreme weather on port operations has become increasingly prominent. Short-term weather will have a great impact on the safety and normal operation of logistics and other enterprises. Extreme weather has caused a great threat to people's property and even life. Timely access to weather forecasts can help governments and industry-related organizations make correct decisions.
- an end-to-end solution can be proposed using deep learning methods to solve this nonlinear complex problem. It extracts low-level features through multi-layer network structure and nonlinear transformation, forms an abstract high-level representation, discovers the probability distribution characteristics of data, and then predicts future precipitation. Since artificial neural networks can model nonlinear systems, using deep neural networks for short-term precipitation prediction has great potential, and the research on this subject has practical significance.
- the realization of the precipitation prediction problem can be seen as predicting the radar echo image after a certain moment through a series of radar echo images before that moment, in a series of inputs before this moment, between the previous input and the next input There is a certain relationship, and they all have different degrees of influence on the prediction results.
- Each output of the feedforward neural network only depends on the current input, and does not consider the mutual influence of inputs at different times, so it is not suitable for dealing with time-space sequence problems. Therefore, RNN is proposed, which is a recurrent neural network. This network is a specialized A network that processes time series data, but it is difficult to deal with too long sequences, that is, it can only have short-term memory and no long-term memory.
- LSTM the long short-term memory network
- This network is a variant of RNN, which combines short-term memory and long-term memory through subtle gate control.
- ConvLSTM is an optimization of LSTM. By adding a convolutional structure in the transition from input to state and state to state, it can better capture the spatiotemporal correlation.
- the SENet module is an attention mechanism, and the fusion of the SENet module can be further improved. The ability of the network to process spatiotemporal correlation information is more suitable for precipitation forecasting.
- the purpose of the present invention is to provide a short-imminent precipitation prediction method based on ConvLSTM and SENet attention mechanism.
- the result expression of this method is more intuitive, the final result is more accurate, and the global features are fully utilized, so that the learned The global information is more reasonable; at the same time, the present invention significantly improves the effect of the network model by adding a SENet attention module under the condition of adding a small number of parameters.
- the present invention does not change the size of the image during the training process, and perfectly preserves the details and edge information of the image.
- the present invention provides a short-term precipitation prediction method, which is realized based on the ConvLSTM network and the SENet attention mechanism module, including the following steps:
- Step 1 preprocessing the meteorological satellite data
- Step 2 radar inversion precipitation, that is, calculate the rain intensity division and cumulative rainfall in the corresponding area through the radar echo;
- Step 3 Build a ConvLSTM network that incorporates the SENet attention module for subsequent model training
- Step 4 importing the aforementioned preprocessed radar echo images into the network for training to obtain a network training model
- Step 5 Use the trained model to generate a predicted precipitation image.
- a further improvement of the present invention is that it also includes step 6, using commonly used precipitation forecasting indicators to evaluate the method.
- the further improvement of the present invention is that the satellite data selects the data of the Fengyun No. 4 satellite, and the radar data selects the radar echo map of the southeastern region of my country.
- a further improvement of the present invention is that in step 2, the so-called radar retrieval precipitation is quantitative precipitation estimation (QPE), which is to calculate the rain intensity distribution and cumulative rainfall in the corresponding area by radar echoes;
- QPE quantitative precipitation estimation
- the existing empirical formula ZR relationship obtains precipitation information, and the ZR relationship reflects the correlation between radar echo and rainfall intensity, that is, the relationship between radar reflection factor Z (unit: mm 6 /m 3 ) and rainfall intensity R (unit mm/h) Satisfies the following formula:
- a and b are empirical constants, a ⁇ 200, and b is between 1.5 and 2, which are determined according to various factors such as different times, different locations, and precipitation types and properties; different rainfall intensities are represented by Different shades of colors are displayed. The darker the color, the greater the rainfall intensity, and vice versa, the less or no precipitation.
- step 3 the precipitation nowcasting can be described as a space-time sequence prediction, in which both the input and the prediction target are time-space sequences;
- ConvLSTM formula is as follows:
- * represents the convolution operation
- a further improvement of the present invention is that the SENet attention module can be inserted into the nonlinear activation function behind each convolutional layer in the ConvLSTM so as to be conveniently integrated into the ConvLSTM network.
- a further improvement of the present invention is to use a precipitation rate threshold of 0.5mm/h (indicating whether it is raining) to convert the predicted and true values into a 0/1 matrix, and calculate hits (successful prediction), misses (failure to predict) and falsealarms (Wrong prediction), the indicators used to evaluate the performance of the model are as follows:
- the present invention also provides a device for implementing the aforementioned short-imminent precipitation prediction method, the device includes at least one computing device, and the computing device includes a memory, a processor, and stored in the memory and A computer program that runs on a processor.
- the method of the present invention is based on ConvLSTM and SENet attention mechanism, and the input and output data are radar echo images, and the result expression is more intuitive.
- a module that can reasonably allocate weights is added to make the final result more accurate.
- the network of the present invention makes full use of the global features, and gives appropriate weights to the long-short-term memory, so that the learned global information is more reasonable.
- the present invention significantly improves the effect of the network model under the condition of adding a small amount of parameters by adding the SENet attention module.
- the present invention does not change the size of the image during the training process, and perfectly preserves the details and edge information of the image.
- Fig. 1 is a block flow diagram of the method of the present invention.
- Fig. 2 is a structural diagram of the ConvLSTM unit in the present invention.
- Figure 3 is the internal structure of ConvLSTM in the present invention.
- Figure 4 is the encoding network and prediction network of ConvLSTM in the present invention.
- Fig. 5 is the basic structure of SENet network in the present invention.
- the present invention uses the Pytorch library in the Python environment on the GeForce GTX 1080Ti processor.
- the present invention uses Adam optimizer and nonlinear activation function ReLU in experiments.
- the number of epochs is set to 100 and the learning rate is set to 0.001.
- the step size was set to 8.
- the short-term precipitation prediction method based on ConvLSTM network and SENet attention mechanism module of the present invention mainly includes the following steps:
- Step 1 preprocessing the meteorological satellite data, the satellite data selects the data of Fengyun No. 4 satellite, and the radar data selects the radar echo map of the southeastern region of my country;
- Step 2 radar inversion precipitation, that is, calculate the rain intensity division and cumulative rainfall in the corresponding area through the radar echo;
- Step 3 Build a ConvLSTM network that incorporates the SENet attention module for subsequent model training
- Step 4 importing the aforementioned preprocessed radar echo images into the network for training to obtain a network training model
- Step 5 Use the trained model to generate a predicted precipitation image
- Step 6 Evaluate the method using commonly used precipitation forecasting indicators.
- the so-called radar inversion of precipitation actually calculates the rain intensity distribution and cumulative rainfall in the corresponding area through radar echoes, which can be called quantitative precipitation estimation in meteorology, or QPE for short.
- Precipitation information is obtained according to the empirical formula ZR relationship between radar echoes and precipitation.
- the ZR relationship reflects the correlation between radar echo and rainfall intensity, that is, the relationship between radar reflection factor Z (unit: mm 6 /m 3 ) and rainfall intensity R (unit: mm/h) satisfies the following formula:
- a and b are empirical constants, a ⁇ 200, and b is between 1.5 and 2, which are determined according to various factors such as different times, different locations, and precipitation types and properties.
- Different rainfall intensities are displayed in different shades of colors on the radar echo map. The darker the color, the greater the rainfall intensity, and vice versa, the smaller or no precipitation.
- step 3 the network structure and principle used and the method of constructing the network are described as follows.
- the precipitation nowcasting problem can be described as a spatio-temporal series forecasting problem, where both the input and the forecast target are spatio-temporal series.
- LSTM By extending the fully connected LSTM, a ConvLSTM with a convolutional structure in both input-to-state and state-to-state transitions is proposed, and the ConvLSTM network can better capture the spatio-temporal correlation.
- ConvLSTM unit structure diagram is shown in Figure 1.
- the formula of ConvLSTM is as follows:
- * represents the convolution operation
- the encoding network and prediction network shown in Figure 3 are formed by stacking multiple ConvLSTM layers.
- the initial state and output of the prediction network are copied from the final state of the encoding network. Because the prediction target has the same dimensionality as the input, all the states in the prediction network are connected and input into a 1*1 convolutional layer to generate the final prediction result.
- the SENet network used in the present invention is a manifestation of the attention mechanism, which allows the network to perform feature recalibration, learn to use global information, selectively emphasize information features, and suppress less useful features.
- the central idea is that for each output channel, predict a constant weight, and perform a weighted operation on each channel to enhance effective information and suppress invalid information.
- the basic structure is shown in Figure 4.
- Ftr represents a convolutional layer that implements a feature map.
- the feature U is passed through a squeeze operation.
- the squeeze operation generates a channel descriptor by aggregating feature maps across spatial dimensions (H*W), thus generating a globally distributed embedding.
- the channel characteristic response of allows information from the global receptive field of the network to be used by all its layers.
- an excitation operation takes the form of a simple pick-and-roll mechanism that takes the embedding as input and produces a set of per-channel modulation weights. These weights are applied to the feature map U, generating the output of the SE block, which can be directly fed into subsequent layers of the network.
- Fsq represents the feature compression according to the spatial dimension, and each two-dimensional feature channel is a one-dimensional vector, which has a global receptive field to some extent, and the output dimension is the same as the number of input feature channels. match. It characterizes the global distribution of responses on feature channels, and enables layers close to the input to gain a global sense.
- Excitation(Fex) represents a mechanism similar to a gate in a recurrent neural network. Weights are produced for each feature channel by a parameter w, which is learned to explicitly model the correlation between feature channels.
- Scale means that the weight of the Excitation output is regarded as the importance of each feature channel after feature selection, and then weighted to the previous features by multiplication channel by channel to complete the original channel dimension. Recalibration of feature maps.
- SENet can be easily integrated into the ConvLSTM network.
- the specific method is to insert the SENet module into the nonlinear activation function behind each convolutional layer in ConvLSTM. It can be achieved by sacrificing a very small calculation cost in exchange for network performance improvement. .
- the indicators used to evaluate the performance of the model are as follows:
- the apparatus for implementing the method of the present invention includes at least one computing device, and the computing device includes a memory, a processor, and a computer program stored in the memory and operable on the processor.
- the computer program is loaded into the processor, the precipitation prediction method based on the ConvLSTM network and the SENet attention module of the present invention is realized.
- the method of the present invention is based on ConvLSTM and SENet attention mechanism, and its input and output data are radar echo images, and the result expression is more intuitive.
- a module that can reasonably allocate weights is added after the activation function of each convolutional layer of ConvLSTM, making the final result more accurate.
- the network of the present invention makes full use of the global features, and gives appropriate weights to the long-short-term memory, so that the learned global information is more reasonable.
- the present invention significantly improves the effect of the network model under the condition of adding a small amount of parameters by adding the SENet attention module.
- the present invention does not change the size of the image during the training process, and perfectly preserves the details and edge information of the image.
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Remote Sensing (AREA)
- Radar, Positioning & Navigation (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- Computer Networks & Wireless Communication (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Biophysics (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Environmental & Geological Engineering (AREA)
- Electromagnetism (AREA)
- Atmospheric Sciences (AREA)
- Biodiversity & Conservation Biology (AREA)
- Ecology (AREA)
- Environmental Sciences (AREA)
- Radar Systems Or Details Thereof (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
Provided in the present invention is an imminent precipitation forecast method, which is implemented on the basis of a ConvLSTM network and an SENet attention mechanism module. The method comprises: pre-processing weather satellite data; performing precipitation inversion by means of a radar; constructing a ConvLSTM network fused with an SENet attention module; importing a pre-processed radar echo image into the network for training, so as to obtain a trained network model; and generating a precipitation forecast image by means of the trained model. In the present invention, both input data and output data are radar echo images, such that result expression is more visual, and a final result is more accurate. In addition, the present invention makes full use of a global feature, such that learned global information is more rational. Furthermore, an SENet attention module is added, such that the effect of a network model is significantly improved. Finally, the size of an image does not change during training in the present invention, and thus details and edge information of the image are perfectly preserved.
Description
本发明属于图像处理技术领域,尤其涉及一种通过雷达回波图像基于ConvLSTM和SENet注意力机制的短临降水预测方法。The invention belongs to the technical field of image processing, and in particular relates to a short-imminent precipitation prediction method based on ConvLSTM and SENet attention mechanism through radar echo images.
智能短临降水预报一般是指六个小时内的降水天气预报,预报精度可以达到公里级和分钟级。和中长期及短期降水预报相比,就需求区域、预报时效等要素而言,对短临预报有着更高的要求。近年来,海洋气候的变化和极端天气事件发生的频率、强度不断增加,极端天气对港口作业的影响日益突显。短临天气对物流等企业的安全及正常运营都会产生较大的影响,极端天气对人们的财产甚至生命都造成了很大威胁。及时获得天气预报,能帮助政府、行业相关组织做出正确决策。目前,针对短临降水预报可以利用深度学习的方法提出一种端到端的方案以解决这种非线性复杂问题。它通过多层的网络结构和非线性变换来提取低层的特征,形成抽象的高层表示,以发现数据的概率分布特征,进而对未来的降水情况进行预测。由于人工神经网络能够对非线性系统进行建模,所以使用深度神经网络进行短临降水预测具有巨大的潜力,课题的研究具有实际意义。Intelligent short-term precipitation forecast generally refers to the precipitation weather forecast within six hours, and the forecast accuracy can reach the kilometer level and minute level. Compared with mid-to-long-term and short-term precipitation forecasts, short-imminent forecasts have higher requirements in terms of demand area, forecast timeliness and other elements. In recent years, changes in marine climate and the frequency and intensity of extreme weather events have been increasing, and the impact of extreme weather on port operations has become increasingly prominent. Short-term weather will have a great impact on the safety and normal operation of logistics and other enterprises. Extreme weather has caused a great threat to people's property and even life. Timely access to weather forecasts can help governments and industry-related organizations make correct decisions. At present, for short-term precipitation forecasting, an end-to-end solution can be proposed using deep learning methods to solve this nonlinear complex problem. It extracts low-level features through multi-layer network structure and nonlinear transformation, forms an abstract high-level representation, discovers the probability distribution characteristics of data, and then predicts future precipitation. Since artificial neural networks can model nonlinear systems, using deep neural networks for short-term precipitation prediction has great potential, and the research on this subject has practical significance.
降水预测问题的实现可以看作通过某个时刻之前的一系列雷达回波图像来预测该时刻之后的雷达回波图像,在该时刻之前的一系列输入中,前一个输入和后一个输入之间是有一定关系的,且对预测结果都有不同程度的影响。前馈神经网络每次的输出都只依赖于当前的 输入,没有考虑不同时刻输入的互相影响,所以不适用于处理时空序列问题,由此提出了RNN即循环神经网络,该网络是一种专门处理时间序列数据的网络,但其难以处理过长的序列,即只能有短期记忆没有长期记忆。为了解决这个问题,LSTM即长短时记忆网络应运而生,该网络是RNN的一种变体,通过精妙的门控制将短期记忆和长期记忆结合起来。ConvLSTM是对LSTM的一种优化,通过在输入到状态和状态到状态的转换中加入卷积结构,可以更好的捕捉时空相关性,SENet模块是一种注意力机制,融合SENet模块可以进一步提高网络对时空相关性信息的处理能力,更适用于降水预测问题。The realization of the precipitation prediction problem can be seen as predicting the radar echo image after a certain moment through a series of radar echo images before that moment, in a series of inputs before this moment, between the previous input and the next input There is a certain relationship, and they all have different degrees of influence on the prediction results. Each output of the feedforward neural network only depends on the current input, and does not consider the mutual influence of inputs at different times, so it is not suitable for dealing with time-space sequence problems. Therefore, RNN is proposed, which is a recurrent neural network. This network is a specialized A network that processes time series data, but it is difficult to deal with too long sequences, that is, it can only have short-term memory and no long-term memory. In order to solve this problem, LSTM, the long short-term memory network, came into being. This network is a variant of RNN, which combines short-term memory and long-term memory through subtle gate control. ConvLSTM is an optimization of LSTM. By adding a convolutional structure in the transition from input to state and state to state, it can better capture the spatiotemporal correlation. The SENet module is an attention mechanism, and the fusion of the SENet module can be further improved. The ability of the network to process spatiotemporal correlation information is more suitable for precipitation forecasting.
发明内容Contents of the invention
本发明的目的是提供一种基于ConvLSTM和SENet注意力机制的短临降水预测方法,采用本方法的结果表达更直观,最后的结果准确性更高,而且充分利用了全局特征,使学习到的全局信息更合理;同时,本发明通过加入SENet注意力模块,在增加少量参数的条件下使网络模型的效果产生显著提升。最后,本发明在训练过程中没有改变图像的尺寸,完美的保留了图像的细节和边缘信息。The purpose of the present invention is to provide a short-imminent precipitation prediction method based on ConvLSTM and SENet attention mechanism. The result expression of this method is more intuitive, the final result is more accurate, and the global features are fully utilized, so that the learned The global information is more reasonable; at the same time, the present invention significantly improves the effect of the network model by adding a SENet attention module under the condition of adding a small number of parameters. Finally, the present invention does not change the size of the image during the training process, and perfectly preserves the details and edge information of the image.
为实现以上目的,本发明提供了一种短临降水预测方法,基于ConvLSTM网络和SENet注意力机制模块而实现,包括以下步骤:In order to achieve the above object, the present invention provides a short-term precipitation prediction method, which is realized based on the ConvLSTM network and the SENet attention mechanism module, including the following steps:
步骤1、对气象卫星数据进行预处理; Step 1, preprocessing the meteorological satellite data;
步骤2、雷达反演降水,即通过雷达回波计算出相应区域的雨强分部和累积雨量;Step 2, radar inversion precipitation, that is, calculate the rain intensity division and cumulative rainfall in the corresponding area through the radar echo;
步骤3、构建融合了SENet注意力模块的ConvLSTM网络,用于后续的模型训练;Step 3. Build a ConvLSTM network that incorporates the SENet attention module for subsequent model training;
步骤4、将前述经过预处理的雷达回波图像导入到网络中进行训练,得到网络训练模型;Step 4, importing the aforementioned preprocessed radar echo images into the network for training to obtain a network training model;
步骤5、使用训练完毕的模型去生成预测降水图像。Step 5. Use the trained model to generate a predicted precipitation image.
本发明的进一步改进在于,还包括步骤6,使用常用的降水预报指标对所述方法进行评价。A further improvement of the present invention is that it also includes step 6, using commonly used precipitation forecasting indicators to evaluate the method.
本发明的进一步改进在于,卫星数据选择风云4号卫星的数据,雷达数据选择我国东南部地区的雷达回波图。The further improvement of the present invention is that the satellite data selects the data of the Fengyun No. 4 satellite, and the radar data selects the radar echo map of the southeastern region of my country.
本发明的进一步改进在于,在步骤2中所谓雷达反演降水为定量降水估计(QPE),其是通过雷达回波计算出相应区域的雨强分布和累积雨量;根据雷达回波和降水之间存在的经验公式Z-R关系获得降水信息,Z-R关系反映了雷达回波和降雨强度的相关性,即雷达反射因子Z(单位:mm
6/m
3)和降雨强度R(单位mm/h)之间满足如下公式:
A further improvement of the present invention is that in step 2, the so-called radar retrieval precipitation is quantitative precipitation estimation (QPE), which is to calculate the rain intensity distribution and cumulative rainfall in the corresponding area by radar echoes; The existing empirical formula ZR relationship obtains precipitation information, and the ZR relationship reflects the correlation between radar echo and rainfall intensity, that is, the relationship between radar reflection factor Z (unit: mm 6 /m 3 ) and rainfall intensity R (unit mm/h) Satisfies the following formula:
Z=aR
b
Z=aR b
其中,a、b是经验常量,a≈200,b在1.5~2之间,根据不同时次、不同地点以及降水类型和性质等多种因素确定;不同的降雨强度在雷达回波图上以深浅不同的颜色表示出来,颜色越深代表降雨强度越大,反之则越小或无降水。Among them, a and b are empirical constants, a≈200, and b is between 1.5 and 2, which are determined according to various factors such as different times, different locations, and precipitation types and properties; different rainfall intensities are represented by Different shades of colors are displayed. The darker the color, the greater the rainfall intensity, and vice versa, the less or no precipitation.
本发明的进一步改进在于,在步骤3中,降水临近预报可以描述为一个时空序列预测,其中输入和预测目标都是时空序列;通过对全连接LSTM进行扩展,提出了在输入到状态和状态到状态转换中都具有卷积结构的ConvLSTM网络,ConvLSTM网络能够捕捉时空相关性;ConvLSTM的公式如下:The further improvement of the present invention is that in step 3, the precipitation nowcasting can be described as a space-time sequence prediction, in which both the input and the prediction target are time-space sequences; ConvLSTM network with convolution structure in state transition, ConvLSTM network can capture space-time correlation; ConvLSTM formula is as follows:
其中,*代表卷积操作,
代表哈达玛运算(对应相乘)。
Among them, * represents the convolution operation, Represents the Hadamard operation (corresponding to multiplication).
本发明的进一步改进在于,所述SENet注意力模块块可以插入到ConvLSTM中的每个卷积层后面的非线性激活函数后从而方便地集成到ConvLSTM网络中。A further improvement of the present invention is that the SENet attention module can be inserted into the nonlinear activation function behind each convolutional layer in the ConvLSTM so as to be conveniently integrated into the ConvLSTM network.
本发明的进一步改进在于,使用0.5mm/h的降水率阈值(表示是否下雨)将预测和真实值转换为0/1矩阵,并计算hits(成功预测)、misses(未能预测)和falsealarms(错误预测),其使用的评价模型的性能的指标如下:A further improvement of the present invention is to use a precipitation rate threshold of 0.5mm/h (indicating whether it is raining) to convert the predicted and true values into a 0/1 matrix, and calculate hits (successful prediction), misses (failure to predict) and falsealarms (Wrong prediction), the indicators used to evaluate the performance of the model are as follows:
为了实现发明目的,本发明还提供了一种装置,用于实施前述的短临降水预测方法,所述装置包括至少一台计算设备,所述计算设备包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序。In order to achieve the object of the invention, the present invention also provides a device for implementing the aforementioned short-imminent precipitation prediction method, the device includes at least one computing device, and the computing device includes a memory, a processor, and stored in the memory and A computer program that runs on a processor.
本发明的有益效果如下:本发明的方法是基于ConvLSTM和SENet注意力机制,其输入和输出的数据均为雷达回波图像,结果表达更直观。同时,在ConvLSTM的每个卷积层的激活函数后面加入 一个能够合理分配权重的模块,使得最后的结果准确性更高。此外,本发明的网络相比传统的LSTM网络,更加充分的利用了全局特征,并且给长短时记忆合适的权重,使学习到的全局信息更合理。再有,本发明通过加入SENet注意力模块,在增加少量参数的条件下使网络模型的效果产生显著提升。最后,本发明在训练过程中没有改变图像的尺寸,完美的保留了图像的细节和边缘信息。The beneficial effects of the present invention are as follows: the method of the present invention is based on ConvLSTM and SENet attention mechanism, and the input and output data are radar echo images, and the result expression is more intuitive. At the same time, after the activation function of each convolutional layer of ConvLSTM, a module that can reasonably allocate weights is added to make the final result more accurate. In addition, compared with the traditional LSTM network, the network of the present invention makes full use of the global features, and gives appropriate weights to the long-short-term memory, so that the learned global information is more reasonable. Furthermore, the present invention significantly improves the effect of the network model under the condition of adding a small amount of parameters by adding the SENet attention module. Finally, the present invention does not change the size of the image during the training process, and perfectly preserves the details and edge information of the image.
图1是本发明方法的流程框图。Fig. 1 is a block flow diagram of the method of the present invention.
图2是本发明中ConvLSTM单元结构图。Fig. 2 is a structural diagram of the ConvLSTM unit in the present invention.
图3是本发明中ConvLSTM内部结构。Figure 3 is the internal structure of ConvLSTM in the present invention.
图4是本发明中ConvLSTM的编码网络和预测网络。Figure 4 is the encoding network and prediction network of ConvLSTM in the present invention.
图5是本发明中SENet网络的基本结构。Fig. 5 is the basic structure of SENet network in the present invention.
为了使本发明的目的、技术方案和优点更加清楚,下面结合附图和具体实施例对本发明进行详细描述。In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments.
需要强调的是,在描述本发明过程中,各种公式和约束条件分别使用前后一致的标号进行区分,但也不排除使用不同的标号标志相同的公式和/或约束条件,这样设置的目的是为了更清楚的说明本发明特征所在。It should be emphasized that in the process of describing the present invention, various formulas and constraints are distinguished by using consistent labels, but it does not exclude the use of different labels to mark the same formulas and/or constraints. The purpose of such setting is In order to illustrate the features of the present invention more clearly.
本发明在GeForce GTX 1080Ti处理器上使用Python环境中的Pytorch库。对于网络优化,本发明在实验中使用了Adam优化器和非线性激活函数ReLU。在网络训练期间,将历次次数设置为100,并将学习率设置为0.001。为了实现精确的收敛,步长设置为8。The present invention uses the Pytorch library in the Python environment on the GeForce GTX 1080Ti processor. For network optimization, the present invention uses Adam optimizer and nonlinear activation function ReLU in experiments. During network training, the number of epochs is set to 100 and the learning rate is set to 0.001. To achieve precise convergence, the step size was set to 8.
如图1所示,本发明的基于ConvLSTM网络和SENet注意力机制 模块的短临降水预测方法,主要包括以下步骤:As shown in Figure 1, the short-term precipitation prediction method based on ConvLSTM network and SENet attention mechanism module of the present invention mainly includes the following steps:
步骤1、对气象卫星数据进行预处理,卫星数据选择风云4号卫星的数据,雷达数据选择我国东南部地区的雷达回波图; Step 1, preprocessing the meteorological satellite data, the satellite data selects the data of Fengyun No. 4 satellite, and the radar data selects the radar echo map of the southeastern region of my country;
步骤2、雷达反演降水,即通过雷达回波计算出相应区域的雨强分部和累积雨量;Step 2, radar inversion precipitation, that is, calculate the rain intensity division and cumulative rainfall in the corresponding area through the radar echo;
步骤3、构建融合了SENet注意力模块的ConvLSTM网络,用于后续的模型训练;Step 3. Build a ConvLSTM network that incorporates the SENet attention module for subsequent model training;
步骤4、将前述经过预处理的雷达回波图像导入到网络中进行训练,得到网络训练模型;Step 4, importing the aforementioned preprocessed radar echo images into the network for training to obtain a network training model;
步骤5、使用训练完毕的模型去生成预测降水图像;Step 5. Use the trained model to generate a predicted precipitation image;
步骤6、使用常用的降水预报指标对方法进行评价。Step 6. Evaluate the method using commonly used precipitation forecasting indicators.
特别地,在步骤2中所用方法的原理和公式说明如下。In particular, the principles and formulas of the method used in step 2 are explained below.
所谓雷达反演降水,实际上是通过雷达回波计算出相应区域的雨强分布和累积雨量,在气象学上可称之为定量降水估计,简称为QPE。根据雷达回波和降水之间存在的经验公式Z-R关系获得降水信息。Z-R关系反映了雷达回波和降雨强度的相关性,即雷达反射因子Z(单位:mm
6/m
3)和降雨强度R(单位mm/h)之间满足如下公式:
The so-called radar inversion of precipitation actually calculates the rain intensity distribution and cumulative rainfall in the corresponding area through radar echoes, which can be called quantitative precipitation estimation in meteorology, or QPE for short. Precipitation information is obtained according to the empirical formula ZR relationship between radar echoes and precipitation. The ZR relationship reflects the correlation between radar echo and rainfall intensity, that is, the relationship between radar reflection factor Z (unit: mm 6 /m 3 ) and rainfall intensity R (unit: mm/h) satisfies the following formula:
Z=aR
b。
Z = aR b .
其中,a、b是经验常量,a≈200,b在1.5~2之间,根据不同时次、不同地点以及降水类型和性质等多种因素确定。不同的降雨强度在雷达回波图上以深浅不同的颜色表示出来,颜色越深代表降雨强度越大,反 之则越小或无降水。Among them, a and b are empirical constants, a≈200, and b is between 1.5 and 2, which are determined according to various factors such as different times, different locations, and precipitation types and properties. Different rainfall intensities are displayed in different shades of colors on the radar echo map. The darker the color, the greater the rainfall intensity, and vice versa, the smaller or no precipitation.
特别地,在步骤3中,其使用的网络结构和原理以及构建网络方法说明如下。降水临近预报问题可以描述为一个时空序列预测问题,其中输入和预测目标都是时空序列。通过对全连接LSTM进行扩展,提出了在输入到状态和状态到状态转换中都具有卷积结构的ConvLSTM,ConvLSTM网络能够更好地捕捉时空相关性。In particular, in step 3, the network structure and principle used and the method of constructing the network are described as follows. The precipitation nowcasting problem can be described as a spatio-temporal series forecasting problem, where both the input and the forecast target are spatio-temporal series. By extending the fully connected LSTM, a ConvLSTM with a convolutional structure in both input-to-state and state-to-state transitions is proposed, and the ConvLSTM network can better capture the spatio-temporal correlation.
ConvLSTM单元结构图如图1所示。ConvLSTM的公式如下:The ConvLSTM unit structure diagram is shown in Figure 1. The formula of ConvLSTM is as follows:
其中,*代表卷积操作;
代表哈达玛运算(对应相乘)。
Among them, * represents the convolution operation; Represents the Hadamard operation (corresponding to multiplication).
ConvLSTM内部结构如图2所示。The internal structure of ConvLSTM is shown in Figure 2.
通过叠加多个ConvLSTM层形成如图3所示的编码网络和预测网络。预测网络的初始状态和输出由编码网络的最后状态复制而来。因为预测目标具有与输入相同的维数,因此将预测网络中的所有状态连接起来,并输入到一个1*1的卷积层中即可生成最终的预测结果。The encoding network and prediction network shown in Figure 3 are formed by stacking multiple ConvLSTM layers. The initial state and output of the prediction network are copied from the final state of the encoding network. Because the prediction target has the same dimensionality as the input, all the states in the prediction network are connected and input into a 1*1 convolutional layer to generate the final prediction result.
注意力机制简单的讲就是把注意力集中在重要的因素上,忽略掉不重要的因素。本发明所用的SENet网络就是注意力机制的一种体现,允许网络执行特征重新校准,学习使用全局信息,选择性的强调信息特征, 抑制掉不太有用的特征,其中心思想是对于每个输出的channel,预测一个常数权重,并对每个channel进行加权运算,增强有效信息,抑制无效信息。基本结构如图4所示。Simply put, the attention mechanism is to focus on important factors and ignore unimportant factors. The SENet network used in the present invention is a manifestation of the attention mechanism, which allows the network to perform feature recalibration, learn to use global information, selectively emphasize information features, and suppress less useful features. The central idea is that for each output channel, predict a constant weight, and perform a weighted operation on each channel to enhance effective information and suppress invalid information. The basic structure is shown in Figure 4.
Ftr表示的是一个卷积层,实现特征映射,特征U通过挤压操作传递,挤压操作通过聚合跨空间维度(H*W)的特征映射生成通道描述符,因而可以产生一个全局分布的嵌入的通道特征响应,允许来自网络的全局接受域的信息被它的所有层使用。之后是激励操作,该操作采用简单的自选门机制的形式,以嵌入作为输入,并产生每个通道调制权值的集合。这些权值被应用到特征映射U中,生成SE块的输出,可以直接输入到网络的后续层中。Ftr represents a convolutional layer that implements a feature map. The feature U is passed through a squeeze operation. The squeeze operation generates a channel descriptor by aggregating feature maps across spatial dimensions (H*W), thus generating a globally distributed embedding. The channel characteristic response of , allows information from the global receptive field of the network to be used by all its layers. This is followed by an excitation operation, which takes the form of a simple pick-and-roll mechanism that takes the embedding as input and produces a set of per-channel modulation weights. These weights are applied to the feature map U, generating the output of the SE block, which can be directly fed into subsequent layers of the network.
Squeeze(Fsq)表示的是按空间维度来进行特征压缩,将每个二维的特征通道一维向量,该向量某种程度上具有全局的感受野,并且输出的维度和输入的特征通道数相匹配。它表征着在特征通道上响应的全局分布,而且使得靠近输入的层也可以获得全局的感受。Squeeze (Fsq) represents the feature compression according to the spatial dimension, and each two-dimensional feature channel is a one-dimensional vector, which has a global receptive field to some extent, and the output dimension is the same as the number of input feature channels. match. It characterizes the global distribution of responses on feature channels, and enables layers close to the input to gain a global sense.
Excitation(Fex)表示的是类似于循环神经网络中门的机制。通过参数w来为每个特征通道生产权重,其中参数w被学习用来显示地建模特征通道间的相关性。Excitation(Fex) represents a mechanism similar to a gate in a recurrent neural network. Weights are produced for each feature channel by a parameter w, which is learned to explicitly model the correlation between feature channels.
Scale(Fscale)表示的是将Excitation的输出的权重看做是进过特征选择后的每个特征通道的重要性,然后通过乘法逐通道加权到先前的特征上,完成在通道维度上的对原始特征图的重标定。Scale (Fscale) means that the weight of the Excitation output is regarded as the importance of each feature channel after feature selection, and then weighted to the previous features by multiplication channel by channel to complete the original channel dimension. Recalibration of feature maps.
SENet可以方便的集成到ConvLSTM网络中,具体方法是将SENet模块插入到ConvLSTM中的每个卷积层后面的非线性激活函数后即可,可以通过牺牲非常小的计算成本来换取网络的性能提升。SENet can be easily integrated into the ConvLSTM network. The specific method is to insert the SENet module into the nonlinear activation function behind each convolutional layer in ConvLSTM. It can be achieved by sacrificing a very small calculation cost in exchange for network performance improvement. .
在本发明的短临降水预测方法中,使用0.5mm/h的降水率阈值(表示是否下雨)将预测和真实值转换为0/1矩阵,并计算hits(成功预测),misses(未能预测)和falsealarms(错误预测),其使用的评价模型的性能的指标如下:In the short-term precipitation prediction method of the present invention, use the precipitation rate threshold of 0.5mm/h (representing whether it is raining) to convert the prediction and the real value into a 0/1 matrix, and calculate hits (successful prediction), misses (failed prediction) and falsealarms (wrong prediction), the indicators used to evaluate the performance of the model are as follows:
实施本发明方法的装置中包括至少一台计算设备,所述计算设备包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序。所述计算机程序被加载至处理器时实现本发明的基于ConvLSTM网络和SENet注意力模块的降水预测方法。The apparatus for implementing the method of the present invention includes at least one computing device, and the computing device includes a memory, a processor, and a computer program stored in the memory and operable on the processor. When the computer program is loaded into the processor, the precipitation prediction method based on the ConvLSTM network and the SENet attention module of the present invention is realized.
本发明的方法是基于ConvLSTM和SENet注意力机制,其输入和输出的数据均为雷达回波图像,结果表达更直观。同时,在ConvLSTM的每个卷积层的激活函数后面加入一个能够合理分配权重的模块,使得最后的结果准确性更高。此外,本发明的网络相比传统的LSTM网络,更加充分的利用了全局特征,并且给长短时记忆合适的权重,使学习到的全局信息更合理。再有,本发明通过加入SENet注意力模块,在增加少量参数的条件下使网络模型的效果产生显著提升。最后,本发明在训练过程中没有改变图像的尺寸,完美的保留了图像的细节和边缘信息。The method of the present invention is based on ConvLSTM and SENet attention mechanism, and its input and output data are radar echo images, and the result expression is more intuitive. At the same time, a module that can reasonably allocate weights is added after the activation function of each convolutional layer of ConvLSTM, making the final result more accurate. In addition, compared with the traditional LSTM network, the network of the present invention makes full use of the global features, and gives appropriate weights to the long-short-term memory, so that the learned global information is more reasonable. Furthermore, the present invention significantly improves the effect of the network model under the condition of adding a small amount of parameters by adding the SENet attention module. Finally, the present invention does not change the size of the image during the training process, and perfectly preserves the details and edge information of the image.
以上实施例仅用以说明本发明的技术方案而非限制,尽管参照较佳实施例对本发明进行了详细说明,本领域的普通技术人员应当理解,可 以对本发明的技术方案进行修改或者等同替换,而不脱离本发明技术方案的精神和范围。The above embodiments are only used to illustrate the technical solutions of the present invention without limitation. Although the present invention has been described in detail with reference to preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced. Without departing from the spirit and scope of the technical solution of the present invention.
Claims (8)
- 一种短临降水预测方法,基于ConvLSTM网络和SENet注意力机制模块而实现,其特征在于,包括以下步骤:A short-term precipitation prediction method is implemented based on the ConvLSTM network and the SENet attention mechanism module, and is characterized in that it includes the following steps:步骤1、对气象卫星数据进行预处理;Step 1, preprocessing the meteorological satellite data;步骤2、雷达反演降水,即通过雷达回波计算出相应区域的雨强分部和累积雨量;Step 2, radar inversion precipitation, that is, calculate the rain intensity division and cumulative rainfall in the corresponding area through the radar echo;步骤3、构建融合了SENet注意力模块的ConvLSTM网络,用于后续的模型训练;Step 3. Build a ConvLSTM network that incorporates the SENet attention module for subsequent model training;步骤4、将前述经过预处理的雷达回波图像导入到网络中进行训练,得到网络训练模型;Step 4, importing the aforementioned preprocessed radar echo images into the network for training to obtain a network training model;步骤5、使用训练完毕的模型去生成预测降水图像。Step 5. Use the trained model to generate a predicted precipitation image.
- 根据权利要求1所述的短临降水预测方法,其特征在于:还包括步骤6,使用常用的降水预报指标对所述方法进行评价。The short-imminent precipitation prediction method according to claim 1, further comprising step 6 of evaluating the method using commonly used precipitation forecasting indicators.
- 根据权利要求2所述的短临降水预测方法,其特征在于:卫星数据选择风云4号卫星的数据,雷达数据选择我国东南部地区的雷达回波图。The short-imminent precipitation prediction method according to claim 2, characterized in that: the satellite data selects the data of Fengyun No. 4 satellite, and the radar data selects the radar echo map of southeastern my country.
- 根据权利要求2所述的短临降水预测方法,其特征在于:在步骤2中所谓雷达反演降水为定量降水估计(QPE),其是通过雷达回波计算出相应区域的雨强分布和累积雨量;根据雷达回波和降水之间存在的经验公式Z-R关系获得降水信息,Z-R关系反映了雷达回波和降雨强度的相关性,即雷达反射因子Z(单位:mm 6/m 3)和降雨强度R(单位mm/h)之间满足如下公式: The method for predicting short-term precipitation according to claim 2, characterized in that: in step 2, the so-called radar retrieval precipitation is quantitative precipitation estimation (QPE), which calculates the distribution and accumulation of rain intensity in the corresponding area by radar echoes Rainfall: Precipitation information is obtained according to the empirical formula ZR relationship between radar echo and precipitation. The ZR relationship reflects the correlation between radar echo and rainfall intensity, that is, radar reflection factor Z (unit: mm 6 /m 3 ) and rainfall The strength R (unit mm/h) satisfies the following formula:Z=aR b Z=aR b其中,a、b是经验常量,a≈200,b在1.5~2之间,根据不同时次、不同地 点以及降水类型和性质等多种因素确定;不同的降雨强度在雷达回波图上以深浅不同的颜色表示出来,颜色越深代表降雨强度越大,反之则越小或无降水。Among them, a and b are empirical constants, a≈200, and b is between 1.5 and 2, which are determined according to various factors such as different times, different locations, and precipitation types and properties; different rainfall intensities are represented by Different shades of colors are displayed. The darker the color, the greater the rainfall intensity, and vice versa, the less or no precipitation.
- 根据权利要求3所述的短临降水预测方法,其特征在于:在步骤3中,降水临近预报可以描述为一个时空序列预测,其中输入和预测目标都是时空序列;通过对全连接LSTM进行扩展,提出了在输入到状态和状态到状态转换中都具有卷积结构的ConvLSTM网络,ConvLSTM网络能够捕捉时空相关性;ConvLSTM的公式如下:The short-imminent precipitation prediction method according to claim 3, characterized in that: in step 3, the precipitation nowcasting can be described as a space-time sequence prediction, wherein the input and the prediction target are all time-space sequences; by extending the full connection LSTM , a ConvLSTM network with a convolutional structure in both input-to-state and state-to-state transitions is proposed. The ConvLSTM network can capture spatiotemporal correlations; the formula of ConvLSTM is as follows:
- 根据权利要求4所述的短临降水预测方法,其特征在于:所述SENet注意力模块块可以插入到ConvLSTM中的每个卷积层后面的非线性激活函数后从而方便地集成到ConvLSTM网络中。The short-term precipitation prediction method according to claim 4, characterized in that: the SENet attention module block can be inserted into the nonlinear activation function behind each convolution layer in ConvLSTM so as to be easily integrated into the ConvLSTM network .
- 根据权利要求5所述的短临降水预测方法,其特征在于:使用0.5mm/h的降水率阈值(表示是否下雨)将预测和真实值转换为0/1矩阵,并计算hits(成功预测)、misses(未能预测)和falsealarms(错误预测),其使用的评价模型的性能的指标如下:The short-term precipitation prediction method according to claim 5, characterized in that: use the precipitation rate threshold of 0.5mm/h (representing whether it rains) to convert prediction and real value into 0/1 matrix, and calculate hits (successful prediction ), misses (failed to predict) and falsealarms (wrong prediction), the indicators used to evaluate the performance of the model are as follows:
- 一种装置,其特征在于:用于实施如权利要求1-7任一项所述的短临降水预测方法,所述装置包括至少一台计算设备,所述计算设备包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序。A device, characterized in that: it is used to implement the short-imminent precipitation prediction method according to any one of claims 1-7, the device includes at least one computing device, and the computing device includes a memory, a processor and a storage A computer program in memory and executable on a processor.
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111498466.5A CN114139690B (en) | 2021-12-09 | 2021-12-09 | Short-term precipitation prediction method and device |
CN202111498466.5 | 2021-12-09 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2023103587A1 true WO2023103587A1 (en) | 2023-06-15 |
Family
ID=80385186
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2022/125568 WO2023103587A1 (en) | 2021-12-09 | 2022-10-17 | Imminent precipitation forecast method and apparatus |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN114139690B (en) |
WO (1) | WO2023103587A1 (en) |
Cited By (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116719002A (en) * | 2023-08-08 | 2023-09-08 | 北京弘象科技有限公司 | Quantitative precipitation estimation method, quantitative precipitation estimation device, electronic equipment and computer storage medium |
CN116739048A (en) * | 2023-08-16 | 2023-09-12 | 合肥工业大学 | Multi-element transducer-based lightning long-term prediction model, method and system |
CN116824372A (en) * | 2023-06-21 | 2023-09-29 | 中国水利水电科学研究院 | Urban rainfall prediction method based on Transformer |
CN116992231A (en) * | 2023-07-18 | 2023-11-03 | 浙江工业大学 | Short-time rainfall prediction method |
CN117094452A (en) * | 2023-10-20 | 2023-11-21 | 浙江天演维真网络科技股份有限公司 | Drought state prediction method, and training method and device of drought state prediction model |
CN117237677A (en) * | 2023-11-15 | 2023-12-15 | 南京信息工程大学 | Precipitation prediction correction method for overall similarity of strong precipitation space based on deep learning |
CN117313823A (en) * | 2023-11-28 | 2023-12-29 | 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) | Mixed distributed parallel training method and system for convolutional neural network |
CN117368881A (en) * | 2023-12-08 | 2024-01-09 | 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) | Multi-source data fusion long-sequence radar image prediction method and system |
CN117909887A (en) * | 2024-03-19 | 2024-04-19 | 国网山东省电力公司嘉祥县供电公司 | Multi-early warning method and system for lightning of power grid |
CN117991412A (en) * | 2024-04-07 | 2024-05-07 | 无锡九方科技有限公司 | Extreme precipitation prediction method and system based on multi-mode data |
CN118050729A (en) * | 2024-04-15 | 2024-05-17 | 南京信息工程大学 | Improved U-Net-based radar echo time downscaling method |
CN118094487A (en) * | 2024-04-26 | 2024-05-28 | 成都信息工程大学 | Method and system for predicting precipitation by multi-source meteorological elements based on space-time perception mechanism |
CN118133671A (en) * | 2024-03-18 | 2024-06-04 | 夏萃慧 | Method for constructing climate influence traceability model |
CN118365475A (en) * | 2024-06-14 | 2024-07-19 | 长扬科技(北京)股份有限公司 | Intelligent monitoring method and device for photovoltaic power plant |
CN118365498A (en) * | 2024-06-17 | 2024-07-19 | 航天宏图信息技术股份有限公司 | Space-time refined solar resource short-term forecasting method, device, equipment and medium |
CN118585960A (en) * | 2024-08-05 | 2024-09-03 | 水利部交通运输部国家能源局南京水利科学研究院 | Satellite-ground precipitation fusion method and system with consideration of dry-wet distribution and strong precipitation landing area |
Families Citing this family (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114139690B (en) * | 2021-12-09 | 2024-07-26 | 南京邮电大学 | Short-term precipitation prediction method and device |
CN114740549B (en) * | 2022-04-12 | 2024-02-06 | 北京墨迹风云科技股份有限公司 | Short-term precipitation weather prediction method and device, electronic equipment and readable medium |
CN115220133B (en) * | 2022-07-19 | 2023-05-26 | 海南大学 | Rainfall prediction method, device and equipment for multiple meteorological elements and storage medium |
CN116451881B (en) * | 2023-06-16 | 2023-08-22 | 南京信息工程大学 | Short-time precipitation prediction method based on MSF-Net network model |
CN118070846A (en) * | 2024-04-18 | 2024-05-24 | 南京信息工程大学 | Short-term precipitation prediction method, device and storage medium |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112446419A (en) * | 2020-10-29 | 2021-03-05 | 中山大学 | Time-space neural network radar echo extrapolation forecasting method based on attention mechanism |
CN113255972A (en) * | 2021-05-10 | 2021-08-13 | 东南大学 | Short-term rainfall prediction method based on Attention mechanism |
WO2021218424A1 (en) * | 2020-04-30 | 2021-11-04 | 江苏科技大学 | Rbf neural network-based method for sea surface wind speed inversion from marine radar image |
CN114139690A (en) * | 2021-12-09 | 2022-03-04 | 南京邮电大学 | Short-term rainfall prediction method and device |
Family Cites Families (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109064507B (en) * | 2018-08-21 | 2021-06-22 | 北京大学深圳研究生院 | Multi-motion-stream deep convolution network model method for video prediction |
CN110363327B (en) * | 2019-06-04 | 2021-09-03 | 东南大学 | ConvLSTM and 3D-CNN-based short rainfall prediction method |
CN111815037B (en) * | 2020-06-24 | 2022-08-26 | 河海大学 | Interpretable short-critical extreme rainfall prediction method based on attention mechanism |
-
2021
- 2021-12-09 CN CN202111498466.5A patent/CN114139690B/en active Active
-
2022
- 2022-10-17 WO PCT/CN2022/125568 patent/WO2023103587A1/en unknown
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2021218424A1 (en) * | 2020-04-30 | 2021-11-04 | 江苏科技大学 | Rbf neural network-based method for sea surface wind speed inversion from marine radar image |
CN112446419A (en) * | 2020-10-29 | 2021-03-05 | 中山大学 | Time-space neural network radar echo extrapolation forecasting method based on attention mechanism |
CN113255972A (en) * | 2021-05-10 | 2021-08-13 | 东南大学 | Short-term rainfall prediction method based on Attention mechanism |
CN114139690A (en) * | 2021-12-09 | 2022-03-04 | 南京邮电大学 | Short-term rainfall prediction method and device |
Cited By (26)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116824372A (en) * | 2023-06-21 | 2023-09-29 | 中国水利水电科学研究院 | Urban rainfall prediction method based on Transformer |
CN116824372B (en) * | 2023-06-21 | 2023-12-08 | 中国水利水电科学研究院 | Urban rainfall prediction method based on Transformer |
CN116992231A (en) * | 2023-07-18 | 2023-11-03 | 浙江工业大学 | Short-time rainfall prediction method |
CN116719002B (en) * | 2023-08-08 | 2023-10-27 | 北京弘象科技有限公司 | Quantitative precipitation estimation method, quantitative precipitation estimation device, electronic equipment and computer storage medium |
CN116719002A (en) * | 2023-08-08 | 2023-09-08 | 北京弘象科技有限公司 | Quantitative precipitation estimation method, quantitative precipitation estimation device, electronic equipment and computer storage medium |
CN116739048A (en) * | 2023-08-16 | 2023-09-12 | 合肥工业大学 | Multi-element transducer-based lightning long-term prediction model, method and system |
CN116739048B (en) * | 2023-08-16 | 2023-10-20 | 合肥工业大学 | Multi-element transducer-based lightning long-term prediction model, method and system |
CN117094452B (en) * | 2023-10-20 | 2024-02-06 | 浙江天演维真网络科技股份有限公司 | Drought state prediction method, and training method and device of drought state prediction model |
CN117094452A (en) * | 2023-10-20 | 2023-11-21 | 浙江天演维真网络科技股份有限公司 | Drought state prediction method, and training method and device of drought state prediction model |
CN117237677B (en) * | 2023-11-15 | 2024-03-26 | 南京信息工程大学 | Precipitation prediction correction method for overall similarity of strong precipitation space based on deep learning |
CN117237677A (en) * | 2023-11-15 | 2023-12-15 | 南京信息工程大学 | Precipitation prediction correction method for overall similarity of strong precipitation space based on deep learning |
CN117313823A (en) * | 2023-11-28 | 2023-12-29 | 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) | Mixed distributed parallel training method and system for convolutional neural network |
CN117313823B (en) * | 2023-11-28 | 2024-04-12 | 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) | Mixed distributed parallel training method and system for convolutional neural network |
CN117368881B (en) * | 2023-12-08 | 2024-03-26 | 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) | Multi-source data fusion long-sequence radar image prediction method and system |
CN117368881A (en) * | 2023-12-08 | 2024-01-09 | 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) | Multi-source data fusion long-sequence radar image prediction method and system |
CN118133671A (en) * | 2024-03-18 | 2024-06-04 | 夏萃慧 | Method for constructing climate influence traceability model |
CN117909887A (en) * | 2024-03-19 | 2024-04-19 | 国网山东省电力公司嘉祥县供电公司 | Multi-early warning method and system for lightning of power grid |
CN117909887B (en) * | 2024-03-19 | 2024-05-31 | 国网山东省电力公司嘉祥县供电公司 | Multi-early warning method and system for lightning of power grid |
CN117991412A (en) * | 2024-04-07 | 2024-05-07 | 无锡九方科技有限公司 | Extreme precipitation prediction method and system based on multi-mode data |
CN117991412B (en) * | 2024-04-07 | 2024-06-04 | 无锡九方科技有限公司 | Extreme precipitation prediction method and system based on multi-mode data |
CN118050729A (en) * | 2024-04-15 | 2024-05-17 | 南京信息工程大学 | Improved U-Net-based radar echo time downscaling method |
CN118094487A (en) * | 2024-04-26 | 2024-05-28 | 成都信息工程大学 | Method and system for predicting precipitation by multi-source meteorological elements based on space-time perception mechanism |
CN118365475A (en) * | 2024-06-14 | 2024-07-19 | 长扬科技(北京)股份有限公司 | Intelligent monitoring method and device for photovoltaic power plant |
CN118365498A (en) * | 2024-06-17 | 2024-07-19 | 航天宏图信息技术股份有限公司 | Space-time refined solar resource short-term forecasting method, device, equipment and medium |
CN118585960A (en) * | 2024-08-05 | 2024-09-03 | 水利部交通运输部国家能源局南京水利科学研究院 | Satellite-ground precipitation fusion method and system with consideration of dry-wet distribution and strong precipitation landing area |
CN118585960B (en) * | 2024-08-05 | 2024-10-11 | 水利部交通运输部国家能源局南京水利科学研究院 | Satellite-ground precipitation fusion method and system with consideration of dry-wet distribution and strong precipitation landing area |
Also Published As
Publication number | Publication date |
---|---|
CN114139690A (en) | 2022-03-04 |
CN114139690B (en) | 2024-07-26 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2023103587A1 (en) | Imminent precipitation forecast method and apparatus | |
Gao et al. | Earthformer: Exploring space-time transformers for earth system forecasting | |
Ma et al. | Transfer learning for long-interval consecutive missing values imputation without external features in air pollution time series | |
Nikoo et al. | Flood-routing modeling with neural network optimized by social-based algorithm | |
CN110991713B (en) | Irregular area flow prediction method based on multi-graph convolution sum GRU | |
CN111639748A (en) | Watershed pollutant flux prediction method based on LSTM-BP space-time combination model | |
CN113537177A (en) | Flood disaster monitoring and disaster situation analysis method based on visual Transformer | |
CN112132149B (en) | Semantic segmentation method and device for remote sensing image | |
Hosseini-Moghari et al. | Drought forecasting using data-driven methods and an evolutionary algorithm | |
CN111462324B (en) | Online spatiotemporal semantic fusion method and system | |
JPWO2006120724A1 (en) | Geographic information system using neural network | |
CN113297972B (en) | Transformer substation equipment defect intelligent analysis method based on data fusion deep learning | |
CN117494034A (en) | Air quality prediction method based on traffic congestion index and multi-source data fusion | |
CN115792853A (en) | Radar echo extrapolation method based on dynamic weight loss | |
CN113610286A (en) | PM accounting for spatio-temporal correlations and meteorological factors2.5Concentration prediction method and device | |
CN114943365A (en) | Rainfall estimation model establishing method fusing multi-source data and rainfall estimation method | |
CN116415730A (en) | Fusion self-attention mechanism time-space deep learning model for predicting water level | |
CN114662791A (en) | Long time sequence pm2.5 prediction method and system based on space-time attention | |
Sahoo et al. | Multi-step ahead urban water demand forecasting using deep learning models | |
CN117496715B (en) | Space-time traffic flow prediction method based on hidden space-time feature mining | |
CN117556312A (en) | Power grid transmission line fault prediction method and system under extreme weather | |
CN117131991A (en) | Urban rainfall prediction method and platform based on hybrid neural network | |
Li et al. | Hydropower generation forecasting via deep neural network | |
CN115456168B (en) | Training method of reinforcement learning model, energy consumption determining method and device | |
CN115630361A (en) | Attention distillation-based federal learning backdoor defense method |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 22903011 Country of ref document: EP Kind code of ref document: A1 |
|
NENP | Non-entry into the national phase |
Ref country code: DE |