Training method and prediction method of artificial intelligence-based precise rainfall prediction model

Abstract

The present application relates to a training method and a prediction method of a rainfall model based on artificial intelligence precision prediction. It mainly includes the steps of preparing a training data set, pre-training and full network training. With the help of the spatial representation ability of spatial reasoning and the precise prediction ability of machine learning, the effective prediction of regional rainfall and the training of regional rainfall prediction model are completed; by introducing spatial information, temperature, humidity, wind speed and wind direction and other environmental factors in the prediction process, and through weighted fusion, the weight adaptive mode, the rainfall prediction model precision is improved and the complexity is optimized, and the rainfall prediction real-time is improved.

Description

[0001] This application is a divisional application of application number 201911104132.8, entitled "A Regional Rainfall Prediction Method Combining Spatial Reasoning and Machine Learning". Technical Field

[0002] This invention relates to spatial reasoning and machine learning, specifically to a training method and a prediction method for an artificial intelligence-based model for accurate rainfall prediction. Background Technology

[0003] Rainfall forecasting has a significant impact on people's social life and production, and its real-time performance and accuracy are crucial. However, the formation and changes of weather systems are influenced by geographical environment and atmospheric motion, and current rainfall forecasting model equations are very complex, making forecasting decisions difficult and of relatively low quality. Summary of the Invention

[0004] The purpose of this invention is to provide a method for accurately predicting rainfall based on artificial intelligence, which can be used for rapid and accurate monitoring of rainfall in local areas.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for accurate rainfall prediction based on artificial intelligence is proposed. It uses spatial reasoning methods and theories to describe the spatial location and climate status information of the area to be monitored. Then, it uses a deep neural network model in machine learning to analyze the information and predict rainfall. The specific implementation steps include: observation data acquisition, observation data preprocessing, observation data preprocessing feature fusion, observation data feature extraction, observation data feature vector temporal fusion, and rainfall prediction.

[0007] The observation data is collected by using n radiosondes distributed at fixed locations to detect the state of the atmosphere, which is used as the observation data; the observation data includes temperature, humidity, whether it is raining, as well as wind speed and wind direction;

[0008] The observation data preprocessing includes preprocessing spatial information, wind direction and speed, temperature value, and humidity value using feature networks to obtain preprocessed feature values; the feature networks include spatial relationship factor feature networks, wind factor feature networks, temperature factor feature networks, and humidity factor feature networks; the resulting preprocessed feature values ​​include spatial relationship factor features, wind factor features, temperature factor features, and humidity factor features.

[0009] The input to the spatial relationship factor feature network is spatial reasoning based on spatial information, (n+2). 2The intersection matrix describes the spatial relationship between n+2 simple regions, which is the spatial relationship matrix. Specifically, from the perspective of spatial division, it is assumed that the effective range of each radiosonde at a fixed location is a simple region, and the rainfall area within the monitoring area and the observation area is also a simple region; its spatial information is the spatial reasoning of simple spatial regions.

[0010] The spatial relationship factor feature network consists of three cascaded convolutional modules 1 with identical structures. Each convolutional module 1 comprises three convolutional layers connected in parallel, with the output processed by Eltwise addition before being input into a Batch Normalization (BN) layer. The kernel sizes of the three convolutional layers are 1×1, 3×3, and 5×5, with padding of 0, 1, and 2, and a stride of 1. The first convolutional module 1 has 4 output channels, the second has 16, and the third has 32. The spatial relationship factor feature network outputs spatial relationship factor features: 32 channels, with a width and height of n+2.

[0011] The wind factor feature network takes wind speed and wind direction matrices as inputs, and is composed of a concatenated concatenated concatenated convolutional modules 2 and 3 concatenated convolutional modules 2 with identical structures. The wind speed and wind direction matrices are generated from the wind speed and wind direction data in the observation data. Their dimensions are the same as those of the spatial relationship matrix. For each element value, the wind speed matrix is ​​the mean of the wind speed and wind direction of the simple region corresponding to the element in the spatial relationship matrix, and the direction matrix is ​​the angle between the mean direction and the vector between the center of the rainfall area and the center of the monitoring area. The concatenated concatenated layer takes wind speed and direction matrices as inputs and is concatenated using the second dimension. The structure and network layer parameter settings of the three convolutional modules 2 with identical structures in the wind factor feature network are consistent with those of the three convolutional modules 1 with identical structures in the spatial relationship factor feature network. The output of the wind factor feature network is wind factor features with 32 channels and n+2 width and height.

[0012] The temperature factor feature network takes a temperature vector as input and consists of two unpooling layers and two identical convolutional modules 3 cascaded alternately. The temperature vector is composed of temperature values ​​from the observation data of each radiosonde, arranged sequentially, and is an n-dimensional row vector. The two unpooling layers expand the input features to n+2 dimensions in both the column and row directions. The convolutional module 3 consists of a convolutional layer and a batch normalization (BN) layer cascaded. The parameters of the convolutional layer are 8 and 32 output channels respectively; the kernel size is 3×3, padding is 1, and stride is 1. The output of the temperature factor feature network is the temperature factor feature, with 32 channels and a width and height of n+2.

[0013] The humidity factor feature network takes a humidity vector as input; the humidity vector is composed of humidity values ​​from the observation data of each radiosonde arranged sequentially; it is an n-dimensional row vector; the structure of the humidity factor feature network is consistent with that of the temperature factor feature network; the output of the humidity factor feature network is a humidity factor feature with 32 channels and n+2 width and height.

[0014] The observation data preprocessing feature fusion uses a merging network structure to fuse the preprocessed feature values ​​and output an observation data descriptor. The merging network structure employs a weighted network structure to learn a weight value for each channel of the input features, and then uses Axpy and Eltwise layers for weighted merging to obtain the observation data descriptor. The weighted network structure is formed by cascading globalpooling layers, fully connected layers, and sigmoid layers; it outputs the weighted weight values ​​for each channel. The observation data descriptor has 32 channels and a width and height of n+2.

[0015] The observation data feature extraction involves passing the observation data descriptor through a multi-level cascaded convolutional network to obtain the observation data feature vector; the multi-level cascaded convolutional network used in the observation data feature extraction is a VGG16 network structure.

[0016] The observation data feature vector time-series fusion generates observation data feature vectors from multiple collected observation data, and inputs them sequentially into the LSTM network to output observation data feature time-series vectors.

[0017] The rainfall prediction process involves feeding the time-series vector of observed data features into a decision unit, which outputs a rainfall prediction result matrix and a rainfall prediction probability matrix. The decision unit consists of a convolutional module and two fully connected layers. The convolutional module contains one convolutional layer and a batch normalization (BN) layer. The two fully connected layers are output connections to the convolutional module, respectively outputting the rainfall prediction result matrix and the rainfall prediction probability matrix. Both the rainfall prediction result matrix and the rainfall prediction probability matrix have dimensions of n+2, representing (n+2)... 2 The rainfall prediction results and probability values ​​for the corresponding element regions of the intersection matrix. Ideally, all network model parameters are preferably trained offline.

[0018] The present invention also provides a training method for an artificial intelligence-based accurate rainfall prediction model, which can be trained to quickly and accurately monitor rainfall in local areas.

[0019] To achieve the other objective mentioned above, the present invention also provides the following technical solution:

[0020] A training method for an artificial intelligence-based accurate rainfall prediction model, characterized by: preparing a training dataset, pre-training, and full network training;

[0021] Prepare the training dataset:

[0022] Acquire historical observation data of the atmospheric state from n radiosondes over a long period of time, and record the corresponding precipitation status;

[0023] For each observation, a spatial relationship matrix, wind speed matrix, wind direction matrix, temperature vector, and humidity vector are generated. Based on the rainfall status of each simple region, a labeled rainfall matrix is ​​generated according to the intersection matrix representation method, with dimensions and width and height both being n+2.

[0024] The aforementioned spatial relationship matrix, wind speed matrix, wind direction matrix, temperature vector, humidity vector, and labeled rainfall matrix constitute a set of training data.

[0025] The spatial relationship matrix is ​​based on spatial reasoning using spatial information, (n+2). 2 The intersection matrix describes the spatial relationship between n+2 simple regions. Specifically, from the perspective of spatial division, it is assumed that the effective range of each fixed radiosonde is a simple region, and the rainfall area within the monitoring area and the observation area is also a simple region. Its spatial information is the spatial reasoning of the simple spatial regions.

[0026] The wind speed matrix and wind direction matrix are generated from the wind speed and wind direction in the observation data; the dimensions are the same as the spatial relationship matrix. For each element value, the wind speed matrix is ​​the mean value of the wind speed and wind direction of the simple region corresponding to the element in the spatial relationship matrix, and the direction matrix is ​​the angle between the mean direction and the vector between the center of the rainfall area and the center of the monitoring area.

[0027] The temperature vector is an n-dimensional row vector composed of the temperature values ​​from the observation data of each radiosonde arranged sequentially.

[0028] The humidity vector is formed by sequentially arranging the humidity values ​​from the observation data of each radiosonde, and is an n-dimensional row vector.

[0029] Historical values ​​of observed data have time-series attributes, meaning that training data also has time-series attributes; the training data set is generated from the historical values ​​of observed data.

[0030] Pre-training:

[0031] Data preprocessing includes using a feature network to preprocess the spatial information, wind direction and speed, temperature value, and humidity value of the training dataset obtained above to obtain preprocessed feature values;

[0032] The feature network includes a spatial relationship factor feature network, a wind factor feature network, a temperature factor feature network, and a humidity factor feature network; its result preprocessing feature values ​​include spatial relationship factor features, wind factor features, temperature factor features, and humidity factor features.

[0033] The spatial relation factor feature network takes a spatial relation matrix as input and outputs spatial relation factor features: it has 32 channels and n+2 width and height. The spatial relation factor feature network is composed of three cascaded convolutional modules 1 with identical structures. Each convolutional module 1 consists of three convolutional layers connected in parallel, and the output is fed into a BN layer after an Eltwise addition operation. The kernel sizes of the three convolutional layers are 1x1, 3x3, and 5x5, respectively, with padding of 0, 1, and 2, and a stride of 1.

[0034] The number of output channels of the convolutional layer in the first convolutional module 1 is set to 4;

[0035] The number of output channels of the convolutional layer in the second convolutional module 1 is set to 16;

[0036] The number of output channels of the convolutional layer in the third convolutional module 1 is set to 32;

[0037] The wind factor feature network takes wind speed and wind direction matrices as inputs and outputs wind factor features: it has 32 channels and n+2 width and height. The wind factor feature network consists of a concatenated concatenated concatenated layer and three convolutional modules 2 with identical structures. The inputs of the concatenated layer are the wind speed and direction matrices, which are concatenated using the second dimension. The structure and network layer parameter settings of the three convolutional modules 2 with identical structures in the wind factor feature network are consistent with those of the three convolutional modules 1 with identical structures in the spatial relationship factor feature network.

[0038] The temperature factor feature network takes a temperature vector as input and outputs temperature factor features: it has 32 channels and n+2 width and height. The temperature factor feature network consists of two unpooling layers and two identical convolutional modules 3 cascaded alternately. The two unpooling layers expand the input features to n+2 dimensions in both the column and row directions. The convolutional module 3 consists of a convolutional layer and a BN layer cascaded together. The parameters of the convolutional layer are: 8 and 32 output channels respectively, 3x3 kernel size, padding of 1, and stride of 1.

[0039] The humidity factor feature network takes a humidity vector as input and outputs humidity factor features: it has 32 channels and n+2 width and height; the humidity factor feature network has the same structure as the temperature factor feature network.

[0040] Data preprocessing feature fusion: The merging network structure is used to fuse the preprocessed feature values ​​and output a data descriptor;

[0041] The merged network structure uses a weighted network structure to learn a weight value for each channel of the input feature, and then uses an Axpy layer and an Eltwise layer to perform weighted merging to obtain a data descriptor;

[0042] The weighted network structure is formed by cascading a global pooling layer, a fully connected layer, and a sigmoid layer, and outputs the weighted weight values ​​of each channel.

[0043] The data descriptor has 32 channels and a width and height of n+2.

[0044] Data feature extraction: Data descriptors are processed through a multi-level cascaded convolutional network to obtain data feature vectors;

[0045] The multi-level cascaded convolutional network used for data feature extraction is the VGG16 network structure;

[0046] After extracting the above data features, the data is fed into a convolutional network and a fully connected layer. The output features are 1 channel and n+2 width and height. The softmax loss function is used to calculate the loss according to the multi-label attributes.

[0047] Specifically, the process involves: performing forward computation using the training data and calculating the loss with the labeled rainfall matrix; using the obtained loss to backpropagate and update the network parameters; and iteratively updating the network parameters on the training set until the loss is less than a given value or the number of iterations is greater than a given value.

[0048] Full network training:

[0049] Setting up a network:

[0050] First, perform the above steps in sequence: data preprocessing, data preprocessing feature fusion, and data feature extraction.

[0051] Then, the data feature vector time sequence fusion is performed: the training data collected multiple times is used to generate training data feature vectors, which are then input into the LSTM network in sequence to output the training data feature time sequence vectors.

[0052] Finally, rainfall prediction is performed: the time-series vector of the training data features is fed into the decision unit, and the rainfall prediction result matrix and the rainfall prediction probability matrix are output.

[0053] The decision-maker consists of a convolutional module and two fully connected layers; the convolutional module contains a convolutional layer and a batch normalization (BN) layer; the two fully connected layers are output connections of the convolutional module, which output the rainfall prediction result matrix and the rainfall prediction probability matrix, respectively.

[0054] The rainfall prediction result matrix and the rainfall prediction probability matrix both have a dimension of n+2, representing (n+2). 2Rainfall prediction results and probability values ​​for the corresponding element regions of the intersection matrix;

[0055] The network built according to the above steps is used as the network initialization value by the pre-training results of the above steps.

[0056] The pre-trained network structure is a learned network structure, while the newly added LSTM network and decision maker in the above steps are unlearned network structures; the learning rate is set to 0.1 times that of the learned network structure and the unlearned network structure.

[0057] The network parameters are iteratively updated on the training set according to the network training process; until the iteration ends, the best performing model is selected to obtain the rainfall prediction model.

[0058] The present invention has the following beneficial effects:

[0059] This invention provides a training method for an artificial intelligence-based accurate rainfall prediction model. It leverages the spatial representation capabilities of spatial reasoning and the precise prediction capabilities of machine learning to effectively predict regional rainfall and train the regional rainfall prediction model. By incorporating environmental factors such as spatial information, temperature, humidity, and wind speed and direction during the prediction process, and through weighted fusion and adaptive weighting, the accuracy and complexity of the rainfall prediction model are improved, enhancing the real-time performance of rainfall prediction. Furthermore, the introduction of an LSTM network, considering longer time series, significantly improves the model's rainfall prediction accuracy. Attached Figure Description

[0060] Figure 1 This is a flowchart of a method for accurately predicting rainfall based on artificial intelligence.

[0061] Figure 2 This is a schematic diagram of the spatial relationship factor feature network connection.

[0062] Figure 3 This is a schematic diagram of the network connection of wind factor characteristics.

[0063] Figure 4 This is a schematic diagram of the temperature factor feature network connection.

[0064] Figure 5 This is a schematic diagram of the connection of the merged network structure.

[0065] Figure 6 This is a schematic diagram of the weighted network connection. Detailed Implementation

[0066] Example 1

[0067] A method for accurate rainfall prediction based on artificial intelligence uses spatial reasoning methods and theories to describe the spatial location and climate state information of the area to be monitored, and then uses a deep neural network model in machine learning to analyze and predict rainfall; the specific implementation steps are as follows: Figure 1 As shown, it includes the following 6 steps: observation data acquisition, observation data preprocessing, observation data preprocessing feature fusion, observation data feature extraction, observation data feature vector temporal fusion, and rainfall prediction;

[0068] Step 1: Observational data acquisition, using n radiosondes distributed at fixed locations to detect the state of the atmosphere as observational data; the observational data includes temperature value, humidity, whether it is raining, as well as wind speed and wind direction;

[0069] Step 2: Preprocessing of observation data, including preprocessing spatial information, wind direction and speed, temperature value, and humidity value using feature networks to obtain preprocessed feature values; wherein the feature networks include spatial relationship factor feature networks, wind factor feature networks, temperature factor feature networks, and humidity factor feature networks; the resulting preprocessed feature values ​​include spatial relationship factor features, wind factor features, temperature factor features, and humidity factor features.

[0070] The above spatial relationship factor feature network input is spatial reasoning based on spatial information, (n+2). 2 The intersection matrix describes the spatial relationship between n+2 simple regions, i.e., the spatial relationship matrix. Specifically, from a spatial division perspective, it is assumed that the effective range of each fixed-location radiosonde is a simple region, and the rainfall areas within the monitoring and observation areas are also simple regions; its spatial information is the spatial inference of simple spatial regions; the spatial relationship factor feature network is as follows: Figure 2 As shown, the system consists of three cascaded convolutional modules 1 with identical structures. Each convolutional module 1 comprises three convolutional layers connected in parallel, with the output processed by Eltwise addition before being input into a Batch Normalization (BN) layer. The kernel sizes of the three convolutional layers are 1x1, 3x3, and 5x5, with padding of 0, 1, and 2, and a stride of 1. The first convolutional module 1 has 4 output channels, the second has 16, and the third has 32. The spatial relation factor feature network outputs spatial relation factor features: 32 channels, with a width and height of n+2.

[0071] The wind factor feature network takes wind speed and wind direction matrices as input, such as Figure 3As shown, the system consists of a concatenated concatenated concatenated layer and three identical convolutional modules 2. The wind speed and wind direction matrices are generated from the wind speed and wind direction data in the observation data. The dimensions are the same as those of the spatial relationship matrix. For each element value, the wind speed matrix is ​​the mean of the wind speed and wind direction of the corresponding simple region in the spatial relationship matrix, and the direction matrix is ​​the angle between the mean direction and the vectors of the center of the rainfall area and the center of the monitoring area. The input of the concatenated layer is the wind speed matrix and the direction matrix. The second dimension is used for concatenation. The structure and network layer parameter settings of the three identical convolutional modules 2 in the wind factor feature network are consistent with those of the three identical convolutional modules 1 in the spatial relationship factor feature network. The output of the wind factor feature network is the wind factor feature, with 32 channels and n+2 width and height.

[0072] Temperature factor feature network input temperature vector; such as Figure 4 As shown, the network consists of two cascaded unpooling layers and two identical convolutional modules 3. The temperature vector is composed of temperature values ​​from the observation data of each radiosonde, arranged sequentially, and is an n-dimensional row vector. The two unpooling layers expand the input features to n+2 dimensions in both the column and row directions. The convolutional module 3 is composed of a convolutional layer and a batch normalization (BN) layer. The parameters of the convolutional layer are 8 and 32, respectively, with a kernel size of 3×3, padding of 1, and stride of 1. The output of the temperature factor feature network is a temperature factor feature with 32 channels and a width and height of n+2.

[0073] The humidity factor feature network takes a humidity vector as input; the humidity vector is composed of humidity values ​​from the observation data of each radiosonde arranged sequentially; it is an n-dimensional row vector; the structure of the humidity factor feature network is consistent with that of the temperature factor feature network; the output of the humidity factor feature network is a humidity factor feature with 32 channels and n+2 width and height.

[0074] Step 3: Preprocessing and feature fusion of observation data. A merging network structure is used to fuse the preprocessed feature values ​​and output the observation data descriptor; the merging network structure is as follows: Figure 5 As shown, a weighted network structure is used to learn a weight value for each channel of the input feature, and then an Axpy layer and an Eltwise layer are used for weighted merging to obtain the observation data descriptor; the above weighted network structure is as follows. Figure 6 As shown, the data is formed by cascading a global pooling layer, a fully connected layer, and a sigmoid layer; the weighted values ​​of each channel are output; the observation data descriptor has 32 channels, and its width and height are both n+2.

[0075] Step 4: Feature extraction of observation data. The observation data descriptor is processed through a multi-level cascaded convolutional network to obtain the observation data feature vector; the multi-level cascaded convolutional network used for observation data feature extraction is a VGG16 network structure.

[0076] Step 5: Temporal fusion of observation data feature vectors. Observation data feature vectors are generated from multiple collections of observation data and sequentially input into the LSTM network to output the temporal vectors of observation data features.

[0077] Step 6: Rainfall prediction. The time-series vector of observed data features is fed into the decision unit, which outputs a rainfall prediction result matrix and a rainfall prediction probability matrix. The decision unit consists of a convolutional module and two fully connected layers. The convolutional module contains one convolutional layer and a batch normalization (BN) layer. The two fully connected layers are the output connections of the convolutional module, respectively outputting the rainfall prediction result matrix and the rainfall prediction probability matrix. Both the rainfall prediction result matrix and the rainfall prediction probability matrix have dimensions of n+2 (n+2). 2 The rainfall prediction results and probability values ​​for the corresponding element regions of the intersection matrix. All network model parameters in this method are preferably trained offline.

[0078] Example 2

[0079] A training method for an artificial intelligence-based accurate rainfall prediction model includes preparing a training dataset, pre-training, and full network training.

[0080] Acquire historical observation data of the atmospheric state from n radiosondes over a long period of time, and record the corresponding precipitation status;

[0081] For each observation data, following the method in Example 1, a spatial relationship matrix, wind speed matrix, wind direction matrix, temperature vector, and humidity vector are generated. Based on the rainfall status of each simple region, a label rainfall matrix is ​​generated according to the intersection matrix representation method, with dimensions of n+2 for both width and height.

[0082] The aforementioned spatial relationship matrix, wind speed matrix, wind direction matrix, temperature vector, humidity vector, and labeled rainfall matrix constitute a set of training data.

[0083] The historical values ​​of its observation data have time-series attributes, meaning that the training data also has time-series attributes; the training data set is generated from the historical values ​​of the observation data.

[0084] In the pre-training stage, the rainfall model in Example 1 is de-LSTM network and decision maker removed. After feature extraction of the observation data in step 4, it is connected to convolutional network and fully connected layer. The output feature is 1 channel and n+2 width and height. The softmax loss function is used to calculate the loss according to the multi-label attributes.

[0085] Specifically, the process involves using the training data for forward computation and calculating the loss with the labeled rainfall matrix; using the obtained loss to backpropagate and update the network parameters; iteratively updating the network parameters on the training set until the loss is less than a given value or the number of iterations is greater than a given value.

[0086] During the full network training phase, the network was built according to the process in Example 1, and the pre-training results were used as the network initialization values.

[0087] The pre-trained network structure is a learned network structure, while the newly added LSTM network and decision maker are unlearned network structures; the learning rate is set to 0.1 times that of the learned network structure.

[0088] The network parameters are iteratively updated on the training set according to the network training process; at the end of the iteration, the best performing model is selected to obtain the rainfall prediction model.

Claims

1. A training method for a regional rainfall prediction model based on spatial reasoning and machine learning, characterized in that: This includes preparing the training dataset, pre-training, and full network training; Prepare the training dataset: Acquire historical observation data of the atmospheric state from n radiosondes over a long period of time, and record the corresponding precipitation status; For each set of observation data, a spatial relationship matrix, wind speed matrix, wind direction matrix, temperature vector, and humidity vector are generated. Based on the rainfall status of each simple region, a labeled rainfall matrix is ​​generated using the intersection matrix representation method, with dimensions and height of [missing information]. n+2 ; The spatial relationship matrix, wind speed matrix, wind direction matrix, temperature vector, humidity vector, and labeled rainfall matrix are a set of training data. The spatial relationship matrix is ​​based on spatial reasoning using spatial information. Intersection matrix, describing n+2 The spatial relationship of a simple region; specifically, from the perspective of spatial division, it is assumed that the effective range of each fixed radiosonde is a simple region, and the rainfall area within the monitoring area and the observation area is also a simple region. Its spatial information is the spatial reasoning of the simple spatial region. The wind speed matrix and wind direction matrix are generated from the wind speed and wind direction in the observation data; the dimensions are the same as the spatial relationship matrix. For each element value, the wind speed matrix is ​​the mean value of the wind speed and wind direction of the simple region corresponding to the element in the spatial relationship matrix, and the direction matrix is ​​the angle between the mean direction and the vector between the center of the rainfall area and the center of the monitoring area. The temperature vector is composed of temperature values ​​from the observation data of each radiosonde arranged sequentially. n 3D row vector; The humidity vector is composed of humidity values ​​from the observation data of each radiosonde, arranged sequentially. n 3D row vector; Historical values ​​of observed data have time-series attributes, meaning that training data also has time-series attributes; the training data set is generated from the historical values ​​of observed data. Pre-training: Data preprocessing includes using a feature network to preprocess the spatial information, wind direction and speed, temperature value, and humidity value of the training dataset obtained above to obtain preprocessed feature values; The feature network includes a spatial relationship factor feature network, a wind factor feature network, a temperature factor feature network, and a humidity factor feature network; its result preprocessing feature values ​​include spatial relationship factor features, wind factor features, temperature factor features, and humidity factor features. The spatial relation factor feature network takes a spatial relation matrix as input and outputs spatial relation factor features: it has 32 channels and its width and height are both [missing information]. n+2 The spatial relationship factor feature network is composed of three convolutional modules 1 with identical structures cascaded together. Each convolutional module 1 consists of three convolutional layers connected in parallel. The output is fed into a BN layer after an Eltwise layer addition operation. The kernel sizes of the three convolutional layers are 1×1, 3×3, and 5×5, respectively, with padding of 0, 1, and 2, and a stride of 1. The number of output channels of the convolutional layer in the first convolutional module 1 is set to 4; The number of output channels of the convolutional layer in the second convolutional module 1 is set to 16; The number of output channels of the convolutional layer in the third convolutional module 1 is set to 32; The wind factor feature network takes wind speed and wind direction matrices as input and outputs wind factor features: it has 32 channels and its width and height are both [missing information]. n+2 The wind factor feature network consists of a concatenated concatenated concatenated layer and three convolutional modules 2 with identical structures. The input to the concatenated layer is the wind speed matrix and the direction matrix, which are concatenated using the second dimension. The structure and network layer parameter settings of the three convolutional modules 2 with identical structures in the wind factor feature network are consistent with those of the three convolutional modules 1 with identical structures in the spatial relationship factor feature network. The temperature factor feature network takes a temperature vector as input and outputs temperature factor features: it has 32 channels and its width and height are both [missing information]. n+ 2 The temperature factor feature network consists of two depooling layers and two identical convolutional modules 3, cascaded alternately. The two depooling layers respectively expand the input features in both the column and row directions. n+2 The convolutional module 3 consists of a convolutional layer and a batch normalization layer cascaded together. The parameters of the convolutional layer are: the number of output channels is 8 and 32 respectively, the kernel size is 3x3, the padding is 1, and the stride is 1. The humidity factor feature network takes a humidity vector as input and outputs humidity factor features: it has 32 channels and its width and height are both [missing information]. n+ 2 The structure of the humidity factor feature network is consistent with that of the temperature factor feature network. Data preprocessing feature fusion: The merging network structure is used to fuse the preprocessed feature values ​​and output a data descriptor; The merged network structure uses a weighted network structure to learn a weight value for each channel of the input feature, and then uses an Axpy layer and an Eltwise layer to perform weighted merging to obtain a data descriptor; The weighted network structure is formed by cascading a global pooling layer, a fully connected layer, and a sigmoid layer, and outputs the weighted weight values ​​of each channel. The data descriptor has 32 channels and its width and height are both [missing information]. n+2 ; Data feature extraction: Data descriptors are processed through a multi-level cascaded convolutional network to obtain data feature vectors; The multi-level cascaded convolutional network used for data feature extraction is the VGG16 network structure; After extracting the above data features, the data is fed into a convolutional network and a fully connected layer. The output features are 1 channel and 1 width and 1 height. n+2 And use the softmax loss function to calculate the loss according to the multi-label attributes; Specifically, the process involves: performing forward computation using the training data and calculating the loss with the labeled rainfall matrix; using the obtained loss to backpropagate and update the network parameters; and iteratively updating the network parameters on the training set until the loss is less than a given value or the number of iterations is greater than a given value. Full network training: Setting up a network: First, perform the above steps in sequence: data preprocessing, data preprocessing feature fusion, and data feature extraction. Then, the data feature vector time sequence fusion is performed: the training data collected multiple times is used to generate training data feature vectors, which are then input into the LSTM network in sequence to output the training data feature time sequence vectors. Finally, rainfall prediction is performed: the time-series vector of the training data features is fed into the decision unit, and the rainfall prediction result matrix and the rainfall prediction probability matrix are output. The decision-maker consists of a convolutional module and two fully connected layers; the convolutional module contains a convolutional layer and a batch normalization (BN) layer; the two fully connected layers are output connections of the convolutional module, which output the rainfall prediction result matrix and the rainfall prediction probability matrix, respectively. The rainfall prediction result matrix and the rainfall prediction probability matrix both have dimensions of [width and height]. n+2 , characterization Rainfall prediction results and probability values ​​for the corresponding element regions of the intersection matrix; The network built according to the above steps is used as the network initialization value by the pre-training results of the above steps. The pre-trained network structure is a learned network structure, while the newly added LSTM network and decision maker in the above steps are unlearned network structures; the learning rate is set to 0.1 times that of the learned network structure and the unlearned network structure. The network parameters are iteratively updated on the training set according to the network training process; until the iteration ends, the best performing model is selected to obtain the rainfall prediction model.

Images