A method to improve the accuracy of short-term precipitation forecast

Abstract

The present disclosure belongs to the field of short-term precipitation forecast, and specifically relates to a method for improving the accuracy of short-term precipitation forecast, including the following methods: collecting continuous radar echo images for weather nowcasting, and converting the radar echo images into tensors; the tensors are superimposed and processed by a four-layer MFI-LSTM network to obtain the network output tensor at the current moment; the network output tensor at the current moment is converted into a corresponding radar echo map; the short-term precipitation forecast information is collected from the newly acquired radar echo map; the MFI-LSTM network is composed of three parts: a multi-scale feature interaction module, a convolutional LSTM gating mechanism, and a memory feature interaction module. On the basis of the LSTM model, the present invention adds two new structures, a multi-scale feature interaction module and a memory feature interaction module, which enhances the interaction of the feature information of the input data, the hidden state, and the memory unit, and can prevent the original LSTM from accumulating prediction errors in the radar echo extrapolation task, resulting in inaccurate prediction of the final radar echo intensity.

Description

Technical Field

[0001] The present invention belongs to the field of short-term precipitation forecasting, and in particular relates to a method for improving the accuracy of short-term precipitation forecasting. Background Art

[0002] Short-term precipitation forecast refers to the forecast of precipitation intensity for a local area in a relatively short period of time in the future (such as 0-2 hours). It can help workers in various industries such as meteorology, disaster prevention and mitigation, agriculture, and transportation to prepare response measures in advance, prevent or reduce social and economic losses caused by severe weather such as heavy precipitation and storms, and has high practical significance.

[0003] At present, meteorological services such as short-term precipitation forecast, lightning nowcast, typhoon disaster monitoring and early warning mainly rely on radar echo extrapolation. Traditional radar echo extrapolation methods are mainly divided into cross-correlation method, centroid tracking method and optical flow-based method. The cross-correlation method can only capture the movement of individual rain clouds, the centroid tracking method is only applicable to echoes with greater intensity, and the optical flow-based method requires optical flow estimation before echo extrapolation. It is difficult to determine the optimal parameters to obtain the best extrapolation effect.

[0004] Recently, the wide application of deep learning in natural language processing, image classification, image segmentation and other tasks has attracted the attention of researchers in the meteorological field. Researchers have tried to introduce deep learning into radar echo extrapolation tasks and proposed many network models, which can be divided into two categories. One is the model based on convolutional neural network, such as ST-CNN, RainNet, Broad-UNet, etc. The other is the model based on recurrent neural network, such as ConvLSTM, PredRNN, PredRNN++, etc.

[0005] These existing methods generally suffer from the problem of inaccurate radar echo intensity prediction due to the lack of contextual semantic feature information interaction and accumulation of prediction errors. Summary of the invention

[0006] In view of the deficiencies of the prior art, the purpose of the present invention is to provide a method for improving the accuracy of short-term precipitation forecast, which solves the problem in the prior art that the radar echo intensity prediction is not accurate enough, affecting the accuracy of short-term precipitation forecast.

[0007] The purpose of this disclosure can be achieved through the following technical solutions:

[0008] A method for improving the accuracy of short-term precipitation forecasting includes the following methods:

[0009] Collect continuous radar echo images for weather nowcasting, and convert radar echo images into tensors;

[0010] The tensor is processed by stacking through a four-layer MFI-LSTM network to obtain the network output tensor at the current moment;

[0011] Convert the current network output tensor into the corresponding radar echo map;

[0012] Collect short-term precipitation forecast information from the newly acquired radar echo map.

[0013] Furthermore, the tensor is processed by stacking through a four-layer MFI-LSTM network to obtain a more accurate tensor, including:

[0014] The MFI-LSTM network of the first layer receives the tensor converted from the input radar echo map at the current moment, and outputs the information and hidden state in the memory unit of the first layer at the current moment;

[0015] The MFI-LSTM network of the 2nd to 4th layers takes the hidden state transmitted from the previous layer as input and outputs the information and hidden state of the memory unit of the current layer at the current moment;

[0016] The hidden state output by the MFI-LSTM network at the 4th layer is passed through the convolution gate to obtain the network output tensor at the current moment.

[0017] Furthermore, the MFI-LSTM network consists of three parts: a multi-scale feature interaction module, a convolutional LSTM gating mechanism and a memory feature interaction module.

[0018] Furthermore, for the moment before the moment t, t-1 indicates the moment before the moment t, the MFI-LSTM network receives the information in the memory unit of the same layer at the previous moment The hidden state of the same layer at the previous moment and the current input X t ; For example, the MFI-LSTM network in the 2nd to 4th layers uses the hidden state of the previous layer at the current moment as input X t ;

[0019] The hidden state of the same layer at the previous moment and the current input X t After the multi-scale feature interaction module, we get and X′ t ; and X′ t Tensors are concatenated along the channel dimension, and the forget gate f, input gate i and output gate o are obtained by Sigmoid activation function operation, and the update g is obtained by tanh activation function operation;

[0020] The memory unit of the same layer at the previous moment Make a Hadamard product with the forget gate f, and add it to the Hadamard product of the input gate i and the update g to get the intermediate memory unit at the current moment;

[0021] After the tanh activation function is operated and the Hadamard product is performed with the output gate o, the intermediate hidden state at the current moment can be obtained. and After the memory feature interaction module is operated, the information output in the memory unit of the current MFI-LSTM network is obtained and hidden state

[0022] Furthermore, the multi-scale feature interaction module performs three 3×3 dilated convolutions on the input tensor X with dilation coefficients of 1, 2, and 5 respectively. t , perform multi-scale feature extraction, then splice the extracted features of three different scales along the channel, and then reduce the number of channels through a 1×1 convolution to obtain the input tensor of multi-scale feature fusion. The tensor is multiplied by 2 after the Sigmoid activation function operation and combined with the hidden state of the same layer from the previous moment. Perform Hadamard product to obtain the input tensor X′ after multi-scale feature interaction t ;

[0023] X′ t After the Sigmoid activation function is operated, it is multiplied by 2 and the Hadamard product is performed with the input tensor of the multi-scale feature fusion obtained previously to obtain the hidden state after the multi-scale feature interaction.

[0024] Furthermore, the memory feature interaction module is obtained by the convolutional LSTM gating mechanism. and The two are first spliced ​​along the channel, and then Splice the concatenated tensor along the channel again to get the concatenated result Split by channel, and get and

[0025] The result after Sigmoid function operation First with The result after the tanh activation function operation Hadamard Reuse and Subtract tensors of the same size whose elements are all 1 The result after the Sigmoid function operation is obtained Then combine it with Adding together the memory units after memory feature interaction The result after Sigmoid function operation and By making a Hadamard product, we can get the hidden state after the memory features interact.

[0026] Beneficial effects of the present disclosure:

[0027] Based on the LSTM model, the present invention adds two new structures, a multi-scale feature interaction module and a memory feature interaction module, which enhances the interaction of feature information among input data, hidden states and memory units, and can enhance the model's ability to capture and learn the contextual semantic relevance of data at previous and subsequent moments, thereby preventing the problem of inaccurate prediction of the final radar echo intensity due to accumulation of prediction errors in the original LSTM in the radar echo extrapolation task. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0029] Figure 1 The network disclosed herein is a stack of four layers of LSTM units;

[0030] Figure 2 It is a comparison bar chart of CSI values ​​of prediction results of MFI-LSTM units with different numbers of stacked layers in the embodiment of the present disclosure;

[0031] Figure 3 This is a bar chart comparing HSS values ​​of prediction results of MFI-LSTM units with different numbers of stacked layers in an embodiment of the present disclosure;

[0032] Figure 4 is a structural diagram of an LSTM unit in an embodiment of the present disclosure;

[0033] Figure 5 is a structural diagram of a multi-scale feature interaction module in an embodiment of the present disclosure;

[0034] Figure 6 is a structural diagram of a memory feature interaction module in an embodiment of the present disclosure;

[0035] Figure 7 is a bar chart comparing CSI values ​​of prediction results of various models in the embodiment of the present disclosure;

[0036] Figure 8It is a bar chart comparing HSS values ​​of prediction results of various models in the embodiments of the present disclosure. DETAILED DESCRIPTION

[0037] The following will be combined with the drawings in the embodiments of the present disclosure to clearly and completely describe the technical solutions in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, rather than all the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present disclosure.

[0038] like Figure 1 As shown in FIG. 1 , the network model proposed in the present invention is composed of four layers of dual-feature interactive LSTM units stacked together. Compared with a single-layer dual-feature interactive LSTM unit, the stacked units can increase the depth of the network and enable it to better capture the learning data features. Figure 2 , Figure 3 The experiment shown in the figure proves that the network obtained by stacking four layers of this unit has the best effect on the radar echo extrapolation task, because too few layers are not conducive to the network's capture and learning of data features, and too many layers will cause the gradient to disappear, making the model prediction performance worse. This network uses time as the flow direction in the horizontal direction and the flow direction of information to the LSTM of different layers in the vertical direction; specifically, at the first moment t=1, the network initializes the hidden state H and memory unit C respectively by generating a tensor with all elements as 0. At each subsequent moment, the information in the memory unit C is transmitted horizontally to the LSTM unit of the same layer at the next moment, and the hidden state H is transmitted horizontally and vertically at the same time.

[0039] Taking time t as an example, the MFI-LSTM unit at the bottom layer (layer 1) of the network receives the tensor X converted from the input radar echo map at the current moment by the built-in Tensor function of the PyTorch deep learning framework. t , the memory unit of the same layer at the previous moment and hidden state The 2nd to 4th layer units of the network take the hidden state passed from the previous layer as input X t , the rest of the operations are the same as the first layer. The hidden state H of the unit output of the highest layer (layer 4) of the network t After passing through the convolution gate, we get the network output tensor O at the current moment t , after the “tensor-image” conversion, the radar echo image corresponding to the output tensor can be obtained.

[0040] like Figure 4As shown in the figure, the dual-feature interaction LSTM unit of the first layer at time t is used as an example for explanation. A single LSTM unit consists of three parts: a multi-scale feature interaction module, a traditional convolutional LSTM gating mechanism, and a memory feature interaction module. The multi-scale feature interaction module extracts multi-scale features from the current input tensor through three different coefficients of dilated convolution kernels, and then performs a Hadamard product operation with the hidden state of the previous moment to perform multi-scale feature fusion, thereby achieving the purpose of obtaining the contextual semantic feature information of the data from the previous hidden state. The model with this module can mine the contextual semantic correlation of the data before and after, better grasp the law of data change, and achieve more accurate prediction. The memory feature interaction module realizes the interaction of the feature information of the intermediate memory unit and the intermediate hidden state output by the LSTM gating unit through two steps of tensor splicing and gating mechanism. The model with this module can enhance the correlation between the hidden state and the intermediate memory unit, prevent the occurrence of model prediction error accumulation due to bad data, and improve the prediction accuracy.

[0041] The MFI-LSTM unit receives the information from the memory unit in the same layer at the previous moment The hidden state of the same layer at the previous moment and the current input X t (2 to 4 layers will be the hidden state of the previous layer at the current moment (such as ) as input X t ). The hidden state of the same layer at the previous moment and the current input After the multi-scale feature interaction module, we get and X′ t . and X′ t Tensors are concatenated along the channel dimension, and the forget gate f, input gate i and output gate o are obtained by Sigmoid activation function operation, and the update g is obtained by tanh activation function operation. Make a Hadamard product with the forget gate f, and add it to the Hadamard product of the input gate i and the update g to get the intermediate memory unit at the current moment.

[0042] After the tanh activation function is operated and the Hadamard product is performed with the output gate o, the intermediate hidden state at the current moment can be obtained. and After the memory feature interaction module is operated, the output of the current MFI-LSTM is obtained

[0043] The multi-scale feature interaction module in the MFI-LSTM unit is as follows Figure 5As shown in the figure, the module first performs three 3×3 dilated convolutions on the input tensor X with dilation factors of 1, 2, and 5 respectively. t , perform multi-scale feature extraction of 3×3, 5×5 and / or 7×7 regions of three different sizes, then concatenate the extracted features of three different scales along the channel, and then reduce the number of channels through a 1×1 convolution to obtain the input tensor of multi-scale feature fusion. The tensor is multiplied by 2 after the Sigmoid activation function operation and combined with the hidden state of the same layer from the previous moment. Perform Hadamard product to obtain the input tensor X′ after multi-scale feature interaction t . X′ t After the Sigmoid activation function is operated, it is multiplied by 2 and the Hadamard product is performed with the input tensor of the multi-scale feature fusion obtained previously to obtain the hidden state after the multi-scale feature interaction.

[0044] The memory feature interaction module in the MFI-LSTM unit is as follows Figure 6 The intermediate memory unit obtained by the traditional convolutional LSTM gating mechanism With the intermediate hidden state The two are first spliced ​​along the channel, and then Splice the concatenated tensor along the channel again to get the concatenated result Split by channel, and get and The result after Sigmoid function operation First with The result after the tanh activation function operation Hadamard Reuse and Subtract tensors of the same size whose elements are all 1 The result after the Sigmoid function operation is obtained Then combine it with Adding together the memory units after memory feature interaction The result after Sigmoid function operation and By making a Hadamard product, we can get the hidden state after the memory features interact.

[0045] Based on the traditional LSTM model, the present invention adds two new structures, a multi-scale feature interaction module and a memory feature interaction module, which enhances the interaction of feature information among input data, hidden state and memory unit, and can prevent the problem of inaccurate prediction of the final radar echo intensity due to accumulation of prediction errors in the original LSTM in the radar echo extrapolation task.

[0046] To prove the effectiveness of the present invention, the LSTM proposed in the present invention is experimentally verified on the collected radar echo dataset and compared with the traditional ConvLSTM, PredRNN and PredRNN++ networks. The radar echo intensity threshold x is set to 20, 30 and 40, the predicted radar echo map is converted into a 0-1 matrix, and the Critical Success Index (CSI) and Heidke Skill Score (HSS) are used to evaluate the experimental results.

[0047] Figure 7 When the radar echo intensity threshold x is set to 20, 30 and 40, the CSI values ​​of the prediction results of each model are shown. It can be seen that when the threshold x is 20, the CSI of the LSTM proposed in the present invention is 4.23% higher than that of the traditional LSTM, 4.85% higher than that of PredRNN, and 3.54% higher than that of PredRNN++; when the threshold x is 30, the CSI is 42.30% higher than that of ConvLSTM, 15.05% higher than that of PredRNN, and 11.88% higher than that of PredRNN++. When the threshold x is 40, the CSI is 103.82% higher than that of the traditional LSTM, 73.55% higher than that of PredRNN, and 49.09% higher than that of PredRNN++.

[0048] Figure 8 When the radar echo intensity threshold x is set to 20, 30 and 40, the HSS values ​​of the prediction results of each model are shown. It can be seen that when the threshold x is 20, the HSS of the MSI-LSTM proposed in the present invention is 2.35% higher than that of the traditional LSTM, 7.64% higher than that of PredRNN, and 3.63% higher than that of PredRNN++. When the threshold x is 30, the HSS is 33.43% higher than that of ConvLSTM, 14.06% higher than that of PredRNN, and 9.92% higher than that of PredRNN++. When the threshold x is 40, the HSS is 92.38% higher than that of the traditional LSTM, 66.54% higher than that of PredRNN, and 43.42% higher than that of PredRNN++.

[0049] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present disclosure. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0050] The above shows and describes the basic principles, main features and advantages of the present disclosure. Those skilled in the art should understand that the present disclosure is not limited by the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present disclosure. Without departing from the spirit and scope of the present disclosure, the present disclosure may have various changes and improvements, and these changes and improvements all fall within the scope of the present disclosure to be protected.

Claims

1. A method for improving the accuracy of short-term precipitation forecast, characterized in that: The following methods are included: Collect continuous radar echo images for weather nowcasting, and convert radar echo images into tensors; The tensor is processed by stacking through a four-layer MFI-LSTM network to obtain the network output tensor at the current moment; Convert the current network output tensor into the corresponding radar echo map; Collect short-term precipitation forecast information from the newly acquired radar echo map; The MFI-LSTM network consists of three parts: a multi-scale feature interaction module, a convolutional LSTM gating mechanism, and a memory feature interaction module; The multi-scale feature interaction module performs three 3×3 dilated convolutions on the input tensor with dilation coefficients of 1, 2, and 5 respectively. , perform multi-scale feature extraction, then splice the extracted features of three different scales along the channel, and then reduce the number of channels through a 1×1 convolution to obtain the input tensor of multi-scale feature fusion. The tensor is multiplied by 2 after the Sigmoid activation function operation and combined with the hidden state of the same layer from the previous moment. Perform Hadamard product to obtain the input tensor after multi-scale feature interaction ; Will After the Sigmoid activation function is operated, it is multiplied by 2 and the Hadamard product is performed with the input tensor of the multi-scale feature fusion obtained previously to obtain the hidden state after the multi-scale feature interaction. ; The memory feature interaction module is obtained by the convolutional LSTM gating mechanism. and , the two are first spliced ​​along the channel, and then Splice the concatenated tensor along the channel again to get the concatenated result ; Split by channel, and get , and ; The result after the Sigmoid function operation is Sigmoid ( ), Sigmoid( ) First, The result after the tanh activation function operation is tanh ( ) is used as Hadamard product to obtain Sigmoid ( )⨀tanh( ), and then use Subtract tensors of the same size whose elements are all 1 The result after the Sigmoid function operation is 1-Sigmoid ( ); and then compare it with Sigmoid ( )⨀tanh( ) can be added to obtain the memory unit after the memory feature interaction ; The result after the Sigmoid function operation is Sigmoid ( )and By making a Hadamard product, we can get the hidden state after the memory features interact. .

2. A method for improving the accuracy of short-term precipitation forecast according to claim 1, characterized in that: The tensor is processed by stacking through a four-layer MFI-LSTM network to obtain a more accurate tensor, including: The MFI-LSTM network of the first layer receives the tensor converted from the input radar echo map at the current moment, and outputs the information and hidden state in the memory unit of the first layer at the current moment; The MFI-LSTM network of the 2nd to 4th layers takes the hidden state transmitted from the previous layer as input and outputs the information and hidden state of the memory unit of the current layer at the current moment; The hidden state output by the MFI-LSTM network at the 4th layer is passed through the convolution gate to obtain the network output tensor at the current moment.

3. The method for improving the accuracy of short-term precipitation forecast according to claim 1, characterized in that: For the moment before time t, t-1 marks the moment before time t, the MFI-LSTM network receives the information in the memory unit of the same layer at the previous moment , the hidden state of the same layer at the previous moment and the current input ; For example, the MFI-LSTM network in the 2nd to 4th layers uses the hidden state of the previous layer at the current moment as input ; The hidden state of the same layer at the previous moment and the current input After the multi-scale feature interaction module, we get and ; and Tensors are concatenated along the channel dimension, and the forget gate f, input gate i and output gate o are obtained by Sigmoid activation function operation, and the update g is obtained by tanh activation function operation; The memory unit of the same layer at the previous moment Make a Hadamard product with the forget gate f, and add it to the Hadamard product of the input gate i and the update g to get the intermediate memory unit at the current moment; After the tanh activation function is operated and the Hadamard product is performed with the output gate o, the intermediate hidden state at the current moment can be obtained. ; and After the memory feature interaction module is operated, the information output in the memory unit of the current MFI-LSTM network is obtained and hidden state .

Images