A large-range daily precipitation spatio-temporal prediction method based on cellular automaton

Abstract

The application provides a large-range daily precipitation spatio-temporal prediction method based on a cellular automaton, and belongs to the technical field of data processing, and specifically comprises the following steps: preprocessing meteorological variable data; constructing a first cellular automaton model based on historical meteorological characteristics; constructing a second cellular automaton model based on historical state conversion; constructing a convolutional long short-term memory network based on position affine transformation, training the convolutional long short-term memory network by using a training set and a verification set, and outputting prediction values of precipitation and GPD quantile probability at a future time; according to a preliminary prediction result, different strategies are respectively adopted for a non-precipitation state, a common precipitation state and an extreme precipitation state, the prediction result of the convolutional long short-term memory network is combined, and the spatial distribution of precipitation of a first future day is calculated; and then, according to a multi-step prediction result of a cellular state, the spatial distribution of precipitation of multiple future days is calculated by using the same strategy. Through the scheme of the application, the prediction accuracy and robustness are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to a large-range daily precipitation spatiotemporal prediction method based on a cellular automaton. BACKGROUND

[0002] At present, a data-driven model models a precipitation process based on a statistical learning method, can mine a complex nonlinear precipitation spatiotemporal pattern from historical observation data, and can predict a variation trend of future precipitation. In comparison, the data-driven model does not need complex initialization conditions, has strong adaptive ability and nonlinear fitting ability, can significantly reduce training and calculation costs while ensuring better prediction accuracy, and has attracted widespread attention in recent years. However, the existing data-driven model is still limited by spatial sparsity of precipitation distribution and spatial heterogeneity of precipitation patterns. Specifically, the spatial sparsity of precipitation distribution means that precipitation values show a significant long-tail distribution, that is, the number of non-precipitation events is much larger than that of precipitation events, and the number of non-extreme precipitation events is much larger than that of extreme precipitation events, which will cause insufficient sensitivity of the model to precipitation events, and thus cause precipitation event false positives and prediction bias. In addition, the spatial heterogeneity of precipitation patterns means that precipitation laws show significant non-stationarity and regional differences, that is, a unified global parameter model is difficult to maintain consistent performance across regions, which greatly increases the complexity of modeling precipitation patterns.

[0003] Therefore, there is an urgent need for a large-range daily precipitation spatiotemporal prediction method based on a cellular automaton with high prediction accuracy and robustness. SUMMARY

[0004] Therefore, the embodiments of the present application provide a large-range daily precipitation spatiotemporal prediction method based on a cellular automaton, which at least partially solves the problem of poor prediction accuracy and robustness in the prior art.

[0005] The embodiments of the present application provide a large-range daily precipitation spatiotemporal prediction method based on a cellular automaton, comprising:

[0006] Step 1, preprocessing meteorological variable data used for daily precipitation spatiotemporal prediction to obtain a training set and a validation set;

[0007] Step 2, constructing a first cellular automaton model based on historical meteorological characteristics to obtain a preliminary prediction result of a cell state;

[0008] Step 3, constructing a second cellular automaton model based on historical state transitions, and using the second cellular automaton model to iterate the preliminary prediction result to obtain a multi-step prediction result of the cell state;

[0009] Step 4: Construct a convolutional long short-term memory network based on positional affine transformation, train it using the training set and validation set, and output the predicted values ​​of precipitation and GPD quantile probability at future time.

[0010] Step 5: Based on the preliminary prediction results, different strategies are adopted for non-precipitation, normal precipitation, and extreme precipitation states. Combined with the prediction results of the convolutional long short-term memory network, the spatial distribution of precipitation for the first day in the future is calculated. Then, based on the multi-step prediction results of the cell state, the spatial distribution of precipitation for the next few days is calculated using the same strategy.

[0011] According to a specific implementation of an embodiment of the present invention, step 1 specifically includes:

[0012] Step 1.1: Divide the meteorological variable data into training sets. And the validation set vs. the training set, where the training set is the validation set. The parameters of the daily precipitation spatiotemporal prediction model are used to update the model, and the validation set VS is used to test the convergence of the model.

[0013] Step 1.2, Calculate the training set Maximum values ​​of various meteorological variables in China and minimum value For meteorological variable data Standardize:

[0014] ;

[0015] in, This represents standardized meteorological variable data;

[0016] Step 1.3: Define three discrete states for each spatial unit based on precipitation data. The categories are: non-precipitation state, normal precipitation state, and extreme precipitation state. The non-precipitation state is characterized by daily precipitation of less than 1 mm / day, which is represented by label 0. The normal precipitation state is characterized by daily precipitation of not less than 1 mm / day and less than the extreme precipitation threshold, which is represented by label 1. The extreme precipitation state is characterized by daily precipitation of not less than the extreme precipitation threshold, which is represented by label 2.

[0017] Step 1.4, based on the discrete states at a certain time and the previous time. Define state transition type The state transition types include all nine combinations of transitioning from one of the three states—no precipitation, ordinary precipitation, and extreme precipitation—to another or remaining unchanged.

[0018] According to a specific implementation of an embodiment of the present invention, step 2 specifically includes:

[0019] Step 2.1 defines the basic elements of the first cellular automaton model as follows: the cellular space is the study region, the cell is a spatial unit, the cell neighborhood is a first-order Moore neighborhood, and the cell state consists of discrete states. The transformation rules are expressed by the spatiotemporal mapping relationship between historical meteorological variables and future cell states;

[0020] Step 2.2: Organize the standardized meteorological variable data into... ,in, The number of spatial units, The length of the time series. The number of types of meteorological variables, This represents the meteorological variable data within the neighborhood of each cell;

[0021] Step 2.3: Random oversampling is used to enhance precipitation state samples. An extreme gradient boosting tree (XGBoost) model is applied to fit the first transformation rule to each cell to predict whether precipitation will occur in the next moment based on historical meteorological variables. The expression for the first transformation rule is:

[0022] ;

[0023] in, The output of the first transformation rule in the first cellular automaton model represents the first transformation rule. individual cells Whether precipitation occurs at any given time, i.e., whether it is in a precipitation state or a non-precipitation state; Indicates the first individual cells Meteorological variable data within the time neighborhood; Indicates the first The extreme gradient boosting tree model fitted by the first transformation rule of each cell is the... A decision tree; This represents the total number of decision trees in the extreme gradient boosting tree model used to fit the first transformation rule;

[0024] Step 2.4: Random oversampling is used to enhance the extreme precipitation state samples. The extreme gradient boosting tree model XGBoost is applied to fit a second transformation rule to each cell to predict whether extreme precipitation will occur in the next moment based on historical meteorological variables. The expression for the second transformation rule is:

[0025] ;

[0026] In the formula, The output of the second transformation rule in the first cellular automaton model represents the... individual cells Whether extreme precipitation occurs under the condition of precipitation, that is, whether it is a normal precipitation state or an extreme precipitation state; Indicates the first The extreme gradient boosting tree model fitted by the second transformation rule of each cell is the first... A decision tree; This represents the total number of decision trees in the extreme gradient boosting tree model used to fit the second transformation rule;

[0027] Step 2.5: Combining the prediction results of the first and second transformation rules, a preliminary prediction result of the cell state of the first cellular automaton model is obtained, wherein the expression of the preliminary prediction result of the cell state is:

[0028] ;

[0029] In the formula, Indicates the first individual cells The cell state predicted by the first cellular automaton model at time t.

[0030] According to a specific implementation of an embodiment of the present invention, step 3 specifically includes:

[0031] Step 3.1 Define the basic elements of the second cellular automaton model as follows: the cellular space is the study region, the cell is a spatial unit, the cell neighborhood is a first-order Moore neighborhood, and the cell state consists of discrete states. The transformation rules are expressed by the spatiotemporal mapping relationship between the state transition type and the future cell state;

[0032] Step 3.2, organize the state transition type data as follows ,in This indicates the number of state transition types within the neighborhood of each cell;

[0033] Step 3.3: Random oversampling is used to enhance precipitation state samples. An extreme gradient boosting tree model is applied to fit the third transformation rule to predict whether precipitation will occur in the next moment based on historical state transformation types. The expression for the third transformation rule is:

[0034] ;

[0035] In the formula, The output of the third transformation rule in the second cellular automaton model represents the... individual cells Whether precipitation occurs at any given time, i.e., whether it is in a precipitation state or a non-precipitation state; Indicates the first individual cells The state transition type within the time neighborhood; Denotes the extreme gradient boosting tree model used to fit the third transformation rule. A decision tree; This represents the total number of decision trees in the extreme gradient boosting tree model used to fit the third transformation rule;

[0036] Step 3.4: Random oversampling is used to enhance the extreme precipitation state samples. An extreme gradient boosting tree model is applied to fit the fourth transformation rule to predict whether extreme precipitation will occur in the next moment based on historical state transition types. The expression for the fourth transformation rule is:

[0037] ;

[0038] In the formula, The output of the fourth transition rule in the second cellular automaton model represents the... individual cells Whether extreme precipitation occurs under the given conditions of precipitation at any given time, i.e., whether it is a normal precipitation state or an extreme precipitation state; Denotes the extreme gradient boosting tree model used to fit the fourth transformation rule. A decision tree; This represents the total number of decision trees in the extreme gradient boosting tree model used to fit the fourth transformation rule;

[0039] Step 3.5: Combining the prediction results of the third and fourth transformation rules, the multi-step prediction result of the cell state of the second cellular automata model is obtained, wherein the expression of the multi-step prediction result of the cell state is:

[0040] ;

[0041] In the formula, Indicates the first individual cells The cell state predicted by the second cellular automaton model at time t.

[0042] According to a specific implementation of an embodiment of the present invention, step 4 specifically includes:

[0043] Step 4.1: Extract extreme precipitation samples from each cell and fit the generalized Pareto distribution parameter (GPD) of the extreme precipitation in the region where each cell is located, where the... GPD probability density function of each cell for:

[0044] ;

[0045] in, Let represent the values ​​of a random variable that follows a GPD distribution, i.e., the _i_th ... The extreme precipitation value for each cell; Indicates the first The shape parameters of the GPD distribution fitted by each cell; Indicates the first The scale parameter of the GPD distribution fitted by each cell; Indicates the first The location parameters of the GPD distribution fitted by each cell;

[0046] Step 4.2: Based on the generalized Pareto distribution parameters fitted to each cell, calculate the quantile probability of extreme precipitation in the region where each cell is located:

[0047] ;

[0048] In the formula, Indicates the first Extreme precipitation values ​​for each cell The corresponding quantile probability;

[0049] Step 4.3: Organize the precipitation data into... ,in This represents the number of samples input to the PA-ConvLSTM convolutional long short-term memory network. This represents the number of days in historical precipitation data; the cell space is defined as the matrix range. For the length of the cell space, For the width of the cell space, PA-ConvLSTM first encodes the spatiotemporal dependency features of historical precipitation data based on a convolutional long short-term memory network:

[0050] ;

[0051] ;

[0052] ;

[0053] ;

[0054] ;

[0055] ;

[0056] ;

[0057] in, Indicates that the input data is in A slice of time; and They represent and The output hidden state after passing through ConvLSTM; , , and Indicates ConvLSTM in Input gates, forget gates, output gates, and cell states at each time step; These are the learnable parameters of PA-ConvLSTM; and These represent the sigmoid and tanh activation functions, respectively. This represents the dot product operator; This represents the convolution operator;

[0058] Step 4.4: Based on the output hidden state of ConvLSTM, PA-ConvLSTM applies a position-based affine transformation module to perceive the spatial differences in the hidden states of different cells, and outputs the spatially corrected hidden state. :

[0059] ;

[0060] In the formula, Represents the scale parameter of the position-based affine transformation module; This represents the offset parameter of the position-based affine transformation module. and All of these are learnable parameters in PA-ConvLSTM;

[0061] Step 4.5, Hidden State Based on Spatial Correction PA-ConvLSTM uses a fully connected layer to predict the quantile probability of precipitation and GPD distribution at the next time step.

[0062] ;

[0063] In the formula, This indicates that PA-CovLSTM is in The predicted rainfall for each cell at that time; This indicates that PA-CovLSTM is in The quantile probability of the GPD distribution of each cell at time t;

[0064] Step 4.6: Based on the predicted and actual values ​​of precipitation and GPD distribution quantile probabilities obtained from PA-CovLSTM in the training set TS, define the loss function. Calculation error:

[0065] ;

[0066] In the formula, This represents the number of samples in the training set TS; and Let represent the true values ​​of precipitation and GPD distribution quantile probability of the j-th sample in the training set TS, respectively. and Let represent the predicted values ​​of precipitation and GPD distribution quantile probability for the j-th sample in the training set TS, respectively. Represents the adjustable hyperparameters of the loss function; L2 norm operator;

[0067] Step 4.7: Based on the backpropagation algorithm, the error calculated by the loss function is passed to each learnable parameter in PA-ConvLSTM, and the learnable parameters are updated based on the Adam optimizer.

[0068] Step 4.8, iterate through steps 4.6 to 4.7 until the error value of the loss function on the validation set VS reaches the minimum value, stop the iteration process, save the learnable parameters at this time, and you will get the trained PA-ConvLSTM.

[0069] According to a specific implementation of the present invention, the step of employing different strategies for non-precipitation states, ordinary precipitation states, and extreme precipitation states includes:

[0070] When the cell state is predicted to be non-precipitation, the precipitation prediction value of that cell is set to zero.

[0071] When the cell state is predicted to be normal precipitation, the precipitation prediction value directly output by PA-ConvLSTM is used.

[0072] When the cell state is predicted to be an extreme precipitation state, the predicted extreme precipitation value of the cell is obtained by using the GPD quantile probability prediction value output by PA-ConvLSTM and combining it with the GPD distribution parameters pre-fitted for the cell.

[0073] According to a specific implementation of an embodiment of the present invention, the step of obtaining the spatial distribution of precipitation over the next few days includes:

[0074] The cell state and its corresponding state transition type predicted by the second cellular automaton model for the next day are used as inputs for daily iterative prediction. Different strategies are applied to the daily prediction results to calculate the spatial distribution of precipitation over the next few days.

[0075] The large-scale daily precipitation spatiotemporal prediction scheme based on cellular automata in this invention includes: Step 1, preprocessing the meteorological variable data used for daily precipitation spatiotemporal prediction to obtain a training set and a validation set; Step 2, constructing a first cellular automaton model based on historical meteorological characteristics to obtain preliminary prediction results of cellular states; Step 3, constructing a second cellular automaton model based on historical state transitions, and using it to iterate the preliminary prediction results to obtain multi-step prediction results of cellular states; Step 4, constructing a convolutional long short-term memory network based on positional affine transformation, training it using the training set and validation set, and outputting predicted values ​​of precipitation and GPD quantile probability for future times; Step 5, based on the preliminary prediction results, adopting different strategies for non-precipitation states, ordinary precipitation states, and extreme precipitation states, and combining the prediction results of the convolutional long short-term memory network to calculate the spatial distribution of precipitation for the first day in the future, and then calculating the spatial distribution of precipitation for the next few days using the same strategy based on the multi-step prediction results of cellular states.

[0076] The beneficial effects of this invention are as follows: By combining existing sample augmentation techniques and two-stage training strategies with cellular automata models, the spatiotemporal evolution of precipitation events is simulated, thereby mitigating the impact of spatial sparsity. Based on the precipitation event simulation results of cellular automata, a convolutional long short-term memory network based on position-wise affine transformation (PA-ConvLSTM) is proposed. By learning the affine transformation coefficients of different local regions, the differences in the spatiotemporal characteristics of precipitation in different local regions are measured, realizing the spatial heterogeneity modeling of precipitation spatiotemporal patterns and improving prediction accuracy and robustness. Attached Figure Description

[0077] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0078] Figure 1 A flowchart illustrating a method for large-scale spatiotemporal prediction of daily precipitation based on cellular automata, provided in an embodiment of the present invention;

[0079] Figure 2 This is a schematic diagram of the model structure and data flow of a large-scale spatiotemporal prediction method for daily precipitation based on cellular automata, provided for an embodiment of the present invention. Detailed Implementation

[0080] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0081] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0082] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this invention, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0083] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. The illustrations only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0084] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0085] This invention provides a method for large-scale spatiotemporal prediction of daily precipitation based on cellular automata, which can be applied to precipitation prediction in meteorological forecasting scenarios.

[0086] See Figure 1 This is a flowchart illustrating a method for large-scale spatiotemporal prediction of daily precipitation based on cellular automata, provided in an embodiment of the present invention. Figure 1and Figure 2 As shown, the method mainly includes the following steps:

[0087] Step 1: Preprocess the meteorological variable data used for spatiotemporal prediction of daily precipitation to obtain the training set and validation set;

[0088] In practice, the meteorological variable data used for spatiotemporal prediction of daily precipitation are preprocessed, and the specific steps include:

[0089] 1.1) Divide the meteorological variable data into training sets. Validation set vs. test set The training set The parameters of the spatiotemporal prediction model for daily precipitation are used to update the parameters. The validation set VS is used to test the convergence of the model. The test set... Used to evaluate the predictive performance of the model.

[0090] 1.2) Calculate the training set Maximum values ​​of various meteorological variables in China and minimum value For meteorological variable data Standardize:

[0091] ;

[0092] In the formula, This represents standardized meteorological variable data.

[0093] 1.3) Define three discrete states based on precipitation data The precipitation conditions are categorized into three types: non-precipitation, normal precipitation, and extreme precipitation. Non-precipitation refers to daily precipitation less than 1 mm / day, denoted by label 0; normal precipitation refers to daily precipitation not less than 1 mm / day and below the extreme precipitation threshold, denoted by label 1; and extreme precipitation refers to daily precipitation not less than the extreme precipitation threshold, denoted by label 2. The extreme precipitation threshold varies across regions and is defined by the 95th percentile of the local daily precipitation distribution.

[0094] 1.4) Discrete states based on a certain time and the previous time. Define state transition type Non-precipitation state Non-precipitation state is represented by label 0; non-precipitation state Normal precipitation is indicated by label 1; non-precipitation is indicated by label 1. Extreme precipitation conditions are indicated by label 2; ordinary precipitation conditions Non-precipitation status is indicated by label 3; ordinary precipitation status Ordinary precipitation status is indicated by label 4; Ordinary precipitation status Extreme precipitation conditions are indicated by label 5; extreme precipitation conditions Non-precipitation conditions are indicated by label 6; extreme precipitation conditions Ordinary precipitation is indicated by label 7; extreme precipitation is indicated by label 7. Extreme precipitation conditions are indicated by label 8.

[0095] Step 2: Construct a first-cell automaton model based on historical meteorological characteristics to obtain preliminary prediction results of cell states;

[0096] In practical implementation, a cellular automaton model (CA-A) based on historical meteorological characteristics is constructed for preliminary prediction of discrete states. The specific steps include:

[0097] 2.1) Define the basic elements of CA-A: the cell space is the study area, the cell is the basic spatial unit, the cell neighborhood is the first-order Moore neighborhood, the cell state is expressed by the above discrete states, and the transformation rule is expressed by the spatiotemporal mapping relationship between historical meteorological variables and future cell states.

[0098] 2.2) Organize the standardized meteorological variable data into ,in This refers to the number of spatial units (the number of cells). The length of the time series. The number of types of meteorological variables, This represents the meteorological variables within the neighborhood of each cell.

[0099] 2.3) Based on the reorganized meteorological data of each spatial cell, the ordinary precipitation state and extreme precipitation state in the training set TS are merged into a single precipitation state. A random oversampling sample augmentation strategy is used to increase the number of precipitation state samples to alleviate the sample imbalance problem. Based on this, assuming that the spatiotemporal evolution of daily precipitation follows a first-order Markov process, the Extreme Gradient Boosting (XGBoost) model is applied to fit the CA-A precipitation state transition rule to each cell. The XGBoost model can predict whether precipitation will occur in each cell at the next time step based on historical meteorological variables. This process is expressed by the following formula:

[0100] ;

[0101] In the formula, The output of the precipitation state transition rule in the CA-A model represents the first... individual cells Whether precipitation occurs at any given time, i.e., whether it is in a precipitation state or a non-precipitation state; Indicates the first individual cells Characteristics of meteorological variables within the time neighborhood; Indicates the first The precipitation state transition rule of the CA-A model fitted by the N cells in the XGBoost model A decision tree; This represents the total number of decision trees in the XGBoost model used to fit the precipitation state transition rules of the CA-A model.

[0102] 2.4) Extract all precipitation state samples from the training set TS, and use a random oversampling sample augmentation strategy to increase the number of extreme precipitation state samples. Based on this, apply the XGBoost model to fit the CA-A extreme precipitation state transition rule for each cell. The XGBoost model can predict whether extreme precipitation will occur in each cell at the next moment based on historical meteorological variables. This process is expressed by the formula:

[0103] ;

[0104] In the formula, The output of the extreme precipitation state transition rule in the CA-A model represents the first... individual cells Whether extreme precipitation occurs at any given moment, i.e., whether it is a normal precipitation state or an extreme precipitation state; Indicates the first The extreme precipitation state transition rule of the CA-A model fitted by the N cells in the XGBoost model A decision tree; This represents the total number of decision trees in the XGBoost model used to fit the extreme precipitation state transition rules of the CA-A model.

[0105] 2.5) By combining the precipitation state transition rules fitted by the two types of XGBoost models and the extreme precipitation state transition rules, the cell state prediction results of CA-A are obtained. This process is expressed by the following formula:

[0106] ;

[0107] In the formula, Indicates the first individual cells The cell state predicted by CA-A at time t.

[0108] Step 3: Construct a second cellular automaton model based on historical state transitions, and use it to iterate the preliminary prediction results to obtain multi-step prediction results of cellular states;

[0109] In practice, a cellular automaton model based on historical state transitions (CA-N) is constructed for multi-step prediction of discrete states. The specific steps include:

[0110] 3.1) Define the basic elements of CA-N: the cell space is the study area, the cell is the basic spatial unit, the cell neighborhood is the first-order Moore neighborhood, the cell state is expressed by the above discrete states, and the transition rule is expressed by the spatiotemporal mapping relationship between the state transition type and the future cell state.

[0111] 3.2) Organize state transition type data as follows ,in This indicates the state transition type within the neighborhood of each cell.

[0112] 3.3) Based on the state transition type data of each spatial cell, the ordinary precipitation state and extreme precipitation state in the training set TS are merged into a single precipitation state. A random oversampling sample augmentation strategy is used to increase the number of precipitation state samples. Based on this, the XGBoost model is applied to fit the precipitation state transition rule of CA-N. The XGBoost model can predict whether precipitation will occur in each cell at the next time step based on historical meteorological variables. This process is expressed by the following formula:

[0113] ;

[0114] In the formula, The output of the precipitation state transition rule in the CA-N model represents the first... individual cells Whether precipitation occurs at any given time, i.e., whether it is in a precipitation state or a non-precipitation state; Indicates the first individual cells The state transition type within the time neighborhood; This represents the first step in the precipitation state transition rule of the CA-N model within the XGBoost model. A decision tree; This represents the total number of decision trees in the XGBoost model used to fit the precipitation state transition rules of the CA-N model.

[0115] 3.4) Extract all precipitation state samples from the training set TS, and use a random oversampling sample augmentation strategy to increase the number of extreme precipitation state samples. Based on this, apply the XGBoost model to fit the extreme precipitation state transition rules of CA-N. The XGBoost model can predict whether extreme precipitation will occur in each cell at the next moment based on historical meteorological variables. This process is expressed by the formula:

[0116] ;

[0117] In the formula, The output of the extreme precipitation state transition rule in the CA-N model represents the first... individual cells Whether extreme precipitation occurs under the given conditions of precipitation at any given time, i.e., whether it is a normal precipitation state or an extreme precipitation state; This represents the XGBoost model's [number]th [item] in the extreme precipitation state transition rule of the CA-N model. A decision tree; This represents the total number of decision trees in the XGBoost model used to fit the extreme precipitation state transition rules of the CA-N model.

[0118] 3.5) By combining the precipitation state transition rules fitted by the two types of XGBoost models and the extreme precipitation state transition rules, the cell state prediction results of CA-N are obtained. This process is expressed by the following formula:

[0119] ;

[0120] In the formula, Indicates the first individual cells The cell state predicted by CA-N at time t.

[0121] Step 4: Construct a convolutional long short-term memory network based on positional affine transformation, train it using the training set and validation set, and output the predicted values ​​of precipitation and GPD quantile probability at future time.

[0122] In practice, a convolutional long short-term memory network (PA-ConvLSTM) based on positional affine transformation is constructed for numerical prediction of precipitation states. The specific steps include:

[0123] 4.1) Extract extreme precipitation samples from each cell and fit the parameters of the Generalized Pareto Distribution (GPD) of extreme precipitation in the region where each cell is located, where the... GPD probability density function of each cell Expressed as a formula:

[0124] ;

[0125] In the formula, Let represent the values ​​of a random variable that follows a GPD distribution, i.e., the _i_th ... The extreme precipitation value for each cell; Indicates the first The shape parameters of the GPD distribution fitted by each cell; Indicates the first The scale parameter of the GPD distribution fitted by each cell; Indicates the first The location parameters of the GPD distribution fitted by each cell.

[0126] 4.2) Based on the GPD distribution parameters fitted to each cell, calculate the quantile probability of extreme precipitation in the region where each cell is located. This process is expressed by the formula:

[0127] ;

[0128] In the formula, Indicates the first Extreme precipitation values ​​for each cell The corresponding quantile probability. In addition, when the cell state is not an extreme precipitation state, in order to ensure the standardized output of the PA-ConvLSTM model, the quantile probability of its GPD distribution is defined as a specific negative value, which can be taken as a reference to the proportion of extreme precipitation state samples in all cell state samples.

[0129] 4.3) Organize the precipitation data into ,in This indicates the number of samples input to the PA-ConvLSTM. This represents the number of days in historical precipitation data; the cell space is defined as the matrix range. is the length of the cell space (representing the number of cells in each row). The width of the cell space (representing the number of cells in each column). PA-ConvLSTM first encodes the spatiotemporal dependency features of historical precipitation data based on a Convolutional Long Short-Term Memory (ConvLSTM) network. This process can be expressed by the following formula:

[0130] ;

[0131] ;

[0132] ;

[0133] ;

[0134] ;

[0135] ;

[0136] ;

[0137] In the formula, Indicates that the input data is in A slice of time; and They represent and The output hidden state after passing through ConvLSTM; , , and Indicates ConvLSTM in Input gates, forget gates, output gates, and cell states at each time step; These are the learnable parameters of PA-ConvLSTM; and These represent the sigmoid and tanh activation functions, respectively. This represents the dot product operator; This represents the convolution operator.

[0138] 4.4) Based on the output hidden state of ConvLSTM, PA-ConvLSTM uses a position-based affine transformation module to perceive the spatial differences in the hidden states of different cells and outputs spatially corrected hidden states. This process enables the modeling of spatial heterogeneity, which can be expressed by the following formula:

[0139] ;

[0140] In the formula, Represents the scale parameter of the position-based affine transformation module; This represents the offset parameter of the position-based affine transformation module. and All of these are learnable parameters in PA-ConvLSTM.

[0141] 4.5) Hidden states based on spatial correction PA-ConvLSTM uses a fully connected layer to predict the quantile probability of precipitation and GPD distribution at the next time step. This process can be expressed by the following formula:

[0142] ;

[0143] In the formula, This indicates that PA-CovLSTM is in The predicted rainfall for each cell at that time; This indicates that PA-CovLSTM is in The quantile probability of the GPD distribution of each cell at time t.

[0144] 4.6) Based on the predicted and actual values ​​of precipitation and GPD distribution quantile probabilities obtained from PA-CovLSTM in the training set TS, define the loss function. The error is calculated using the following formula:

[0145] ;

[0146] In the formula, This represents the number of samples in the training set TS; and Let represent the true values ​​of precipitation and GPD distribution quantile probability of the j-th sample in the training set TS, respectively. and Let represent the predicted values ​​of precipitation and GPD distribution quantile probability for the j-th sample in the training set TS, respectively. Represents the adjustable hyperparameters of the loss function; This represents the L2 norm operator.

[0147] 4.7) The error calculated by the loss function is passed to each learnable parameter in PA-ConvLSTM based on the backpropagation algorithm, and these parameters are updated based on the Adam optimizer.

[0148] 4.8) Iterate through steps 4.6) to 4.7) until the error value of the loss function on the validation set VS reaches the minimum value, then stop the iteration process, save the learnable parameters at this time, and you will get the trained PA-ConvLSTM.

[0149] Step 5: Based on the preliminary prediction results, different strategies are adopted for non-precipitation, normal precipitation, and extreme precipitation states. Combined with the prediction results of the convolutional long short-term memory network, the spatial distribution of precipitation for the first day in the future is calculated. Then, based on the multi-step prediction results of the cell state, the spatial distribution of precipitation for the next few days is calculated using the same strategy.

[0150] In practice, the spatiotemporal distribution of large-scale precipitation over the next few days is output by combining cellular automata models and PA-ConvLSTM. The specific steps are as follows:

[0151] 5.1) Based on the CA-A model, preliminary predictions of cell states are made. The preliminary predictions of cell states are then analyzed using a "divide and conquer" approach to obtain the spatiotemporal distribution of precipitation over a large area for the next day. Specifically, if the cell state is non-precipitation, the precipitation value is assigned to 0; if the cell state is normal precipitation, the precipitation prediction result from PA-ConvLSTM is used; if the cell state is extreme precipitation, the GPD distribution quantile probability prediction result from PA-ConvLSTM is used to solve for the GPD distribution quantile, which is then used as the precipitation prediction result. This process is expressed by the following formula:

[0152] ;

[0153] In the formula, Indicates the first individual cells Precipitation forecast for the current time; Indicates the first individual cells The cell state predicted by CA-A at time 1; Indicates the first individual cells Precipitation forecasts from PA-ConvLSTM at specific times; Indicates the first individual cells The predicted quantile probability of the GPD distribution at time PA-ConvLSTM.

[0154] 5.2) Based on the preliminary prediction results of the cell states output by the CA-A model, the state transition types are constructed as inputs to the CA-N model. Then, based on the CA-N model, daily iterative predictions of the cell states are performed. Using the same "divide and conquer" approach as in 5.1), the spatiotemporal distribution of large-scale precipitation over the next few days can be obtained. This process is expressed by the following formula:

[0155] ;

[0156] In the formula, Indicates the first individual cells The cell state predicted by CA-N at time 1.

[0157] The large-scale daily precipitation spatiotemporal prediction method based on cellular automata provided in this embodiment combines existing sample augmentation techniques and a two-stage training strategy with a cellular automata model to simulate the spatiotemporal evolution of precipitation events, thereby mitigating the impact of spatial sparsity. Based on the precipitation event simulation results of cellular automata, a Convolutional Long Short-Term Memory Network (PA-ConvLSTM) based on position-wise affine transformation is proposed. By learning the affine transformation coefficients of different local regions, the differences in the spatiotemporal characteristics of precipitation in different local regions are measured, realizing the spatial heterogeneity modeling of precipitation spatiotemporal patterns and improving prediction accuracy and robustness.

[0158] Compared with existing spatiotemporal prediction methods for daily precipitation, this invention has the following two main technical advantages:

[0159] (1) Existing spatiotemporal prediction methods for daily precipitation mainly focus on numerical imbalance when modeling the spatial sparsity of precipitation distribution, neglecting the spatiotemporal evolution of sparse precipitation events, which often leads to overcorrection of the model. This invention combines existing technology with cellular automata models, which can effectively simulate the spatiotemporal evolution of sparse precipitation events and improve the reliability of large-scale daily precipitation prediction results;

[0160] (2) Existing spatiotemporal prediction methods for daily precipitation often ignore the spatial heterogeneity of precipitation patterns, leading to the model's inapplicability in local areas. This invention proposes a convolutional long short-term memory network based on positional affine transformation, which enhances the model's ability to model spatial heterogeneity on the basis of existing technologies and can effectively reduce the prediction bias of daily precipitation over a large area.

[0161] The method of the present invention will be further described below with reference to a specific embodiment, using the LRA5 dataset from the ChaosBench public dataset as the real-world dataset for this embodiment. The LRA5 dataset originates from the ERA5-Land dataset published by a local medium-range weather forecasting center and has undergone post-processing. The LRA5 dataset covers the global land area and spans from January 1, 1979 to December 31, 2023 (this embodiment uses data from 1998 to 2018 for experimental analysis). It has a spatial resolution of 1.5° latitude and longitude and a daily temporal resolution, including 43 meteorological variables such as total precipitation, 2m temperature, 10m wind speed, and total evaporation. The implementation process of the present invention will be described below.

[0162] Based on the test set ES, this invention is compared and analyzed with existing spatiotemporal prediction methods for daily precipitation to verify the effectiveness of this invention. The comparison methods include: ① Gated Recurrent Unit (GRU) model; ② Trajectory Gated Recurrent Unit (TrajGRU) model; ③ Convolutional Long Short-Term Memory (ConvLSTM) network; ④ Hierarchical Two-Fold model; ⑤ Fourcast prediction neural network. Specific steps include:

[0163] The mean absolute error (MAE), root mean square error (RMSE), and precipitation state critical success index (MRSE) are used to measure precipitation conditions. ) and the critical success index for extreme precipitation conditions ( As an evaluation metric, the prediction results of each model are compared using the following formula:

[0164] ;

[0165] ;

[0166] ;

[0167] ;

[0168] In the formula, The number of samples in the test set ES; and Let represent the actual and predicted precipitation values ​​for the j-th sample in the test set ES, respectively. and These represent the number of correctly predicted precipitation and extreme precipitation conditions, respectively. and These represent the number of samples of non-precipitation and non-extreme precipitation states that were incorrectly predicted, respectively. and These represent the number of incorrectly predicted precipitation and extreme precipitation conditions, respectively; smaller MAE and RMSE indicate better prediction performance. and The larger the value, the better the prediction effect.

[0169] Table 1 shows the prediction accuracy of the present invention and the comparative method in the spatiotemporal prediction task of global daily precipitation in the next day.

[0170] Table 1

[0171]

[0172] By comparing the evaluation metrics of the present invention and the comparative methods on the task of global daily precipitation spatiotemporal prediction in the future, the present method outperforms the other five comparative methods in all four evaluation metrics.

[0173] Table 2 shows the prediction accuracy of the present invention and the comparative method in the spatiotemporal prediction of global daily precipitation over the next three days.

[0174] Table 2

[0175]

[0176] By comparing the evaluation metrics of the present invention and the comparative methods in the three-day global daily precipitation spatiotemporal prediction task, the present method outperforms the other five comparative methods in all four evaluation metrics.

[0177] As can be seen from Tables 1 and 2, the method of the present invention can achieve leading prediction results when applied to large-scale daily precipitation spatiotemporal prediction scenarios, enabling more accurate forecasts of precipitation events and extreme precipitation events, and serving important applications such as early warning and emergency response to heavy precipitation disasters, and regional water resources planning and management.

[0178] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof.

[0179] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for large-scale spatiotemporal prediction of daily precipitation based on cellular automata, characterized in that, include: Step 1: Preprocess the meteorological variable data used for spatiotemporal prediction of daily precipitation to obtain the training set and validation set; Step 1 specifically includes: Step 1.1: Divide the meteorological variable data into training sets. And the validation set vs. the training set, where the training set is the validation set. The parameters of the daily precipitation spatiotemporal prediction model are used to update the model, and the validation set VS is used to test the convergence of the model. Step 1.2, Calculate the training set Maximum values ​​of various meteorological variables in China and minimum value For meteorological variable data Standardize: in, This represents standardized meteorological variable data; Step 1.3: Define three discrete states for each spatial unit based on precipitation data. The categories are: non-precipitation state, normal precipitation state, and extreme precipitation state. The non-precipitation state is characterized by daily precipitation of less than 1 mm / day, which is represented by label 0. The normal precipitation state is characterized by daily precipitation of not less than 1 mm / day and less than the extreme precipitation threshold, which is represented by label 1. The extreme precipitation state is characterized by daily precipitation of not less than the extreme precipitation threshold, which is represented by label 2. Step 1.4, based on the discrete states at a certain time and the previous time. Define state transition type The state transition types include all nine combinations of transitioning from one of the three states—no precipitation, ordinary precipitation, and extreme precipitation—to another state or remaining unchanged. Step 2: Construct a first-cell automaton model based on historical meteorological characteristics to obtain preliminary prediction results of cell states; Step 2 specifically includes: Step 2.1 defines the basic elements of the first cellular automaton model as follows: the cellular space is the study region, the cell is a spatial unit, the cell neighborhood is a first-order Moore neighborhood, and the cell state consists of discrete states. The transformation rules are expressed by the spatiotemporal mapping relationship between historical meteorological variables and future cell states; Step 2.2: Organize the standardized meteorological variable data into... ,in, The number of spatial units, The length of the time series. The number of types of meteorological variables, This represents the meteorological variable data within the neighborhood of each cell; Step 2.3: Random oversampling is used to enhance precipitation state samples. An extreme gradient boosting tree (XGBoost) model is applied to fit the first transformation rule to each cell to predict whether precipitation will occur in the next moment based on historical meteorological variables. The expression for the first transformation rule is: in, The output of the first transformation rule in the first cellular automaton model represents the first transformation rule. individual cells Whether precipitation occurs at any given time, i.e., whether it is in a precipitation state or a non-precipitation state; Indicates the first individual cells Meteorological variable data within the time neighborhood; Indicates the first The extreme gradient boosting tree model fitted by the first transformation rule of each cell is the... A decision tree; This represents the total number of decision trees in the extreme gradient boosting tree model used to fit the first transformation rule; Step 2.4: Random oversampling is used to enhance the extreme precipitation state samples. The extreme gradient boosting tree model XGBoost is applied to fit a second transformation rule to each cell to predict whether extreme precipitation will occur in the next moment based on historical meteorological variables. The expression for the second transformation rule is: In the formula, The output of the second transformation rule in the first cellular automaton model represents the... individual cells Whether extreme precipitation occurs under the condition of precipitation, that is, whether it is a normal precipitation state or an extreme precipitation state; Indicates the first The extreme gradient boosting tree model fitted by the second transformation rule of each cell is the first... A decision tree; This represents the total number of decision trees in the extreme gradient boosting tree model used to fit the second transformation rule; Step 2.5: Combining the prediction results of the first and second transformation rules, a preliminary prediction result of the cell state of the first cellular automaton model is obtained, wherein the expression of the preliminary prediction result of the cell state is: In the formula, Indicates the first individual cells The cell state predicted by the first cellular automaton model at time step 1; Step 3: Construct a second cellular automaton model based on historical state transitions, and use it to iterate the preliminary prediction results to obtain multi-step prediction results of cellular states; Step 4: Construct a convolutional long short-term memory network based on positional affine transformation, train it using the training set and validation set, and output the predicted values ​​of precipitation and GPD quantile probability at future time. Step 5: Based on the preliminary prediction results, different strategies are adopted for non-precipitation, normal precipitation, and extreme precipitation states. Combined with the prediction results of the convolutional long short-term memory network, the spatial distribution of precipitation for the first day in the future is calculated. Then, based on the multi-step prediction results of the cell state, the spatial distribution of precipitation for the next few days is calculated using the same strategy.

2. The method according to claim 1, characterized in that, Step 3 specifically includes: Step 3.1 Define the basic elements of the second cellular automaton model as follows: the cellular space is the study region, the cell is a spatial unit, the cell neighborhood is a first-order Moore neighborhood, and the cell state consists of discrete states. The transformation rules are expressed by the spatiotemporal mapping relationship between the state transition type and the future cell state; Step 3.2, organize the state transition type data as follows ,in This indicates the number of state transition types within the neighborhood of each cell; Step 3.3: Random oversampling is used to enhance precipitation state samples. An extreme gradient boosting tree model is applied to fit the third transformation rule to predict whether precipitation will occur in the next moment based on historical state transformation types. The expression for the third transformation rule is: In the formula, The output of the third transformation rule in the second cellular automaton model represents the... individual cells Whether precipitation occurs at any given time, i.e., whether it is in a precipitation state or a non-precipitation state; Indicates the first individual cells The state transition type within the time neighborhood; Denotes the extreme gradient boosting tree model used to fit the third transformation rule. A decision tree; This represents the total number of decision trees in the extreme gradient boosting tree model used to fit the third transformation rule; Step 3.4: Random oversampling is used to enhance the extreme precipitation state samples. An extreme gradient boosting tree model is applied to fit the fourth transformation rule to predict whether extreme precipitation will occur in the next moment based on historical state transition types. The expression for the fourth transformation rule is: In the formula, The output of the fourth transition rule in the second cellular automaton model represents the... individual cells Whether extreme precipitation occurs under the condition of precipitation, that is, whether it is a normal precipitation state or an extreme precipitation state; Denotes the extreme gradient boosting tree model used to fit the fourth transformation rule. A decision tree; This represents the total number of decision trees in the extreme gradient boosting tree model used to fit the fourth transformation rule; Step 3.5: Combining the prediction results of the third and fourth transformation rules, the multi-step prediction result of the cell state of the second cellular automata model is obtained, wherein the expression of the multi-step prediction result of the cell state is: In the formula, Indicates the first individual cells The cell state predicted by the second cellular automaton model at time t.

3. The method according to claim 2, characterized in that, Step 4 specifically includes: Step 4.1: Extract extreme precipitation samples from each cell and fit the generalized Pareto distribution parameter (GPD) of extreme precipitation in the region where each cell is located, where the... GPD probability density function of each cell for: in, Let represent the values ​​of a random variable that follows a GPD distribution, i.e., the _i_th ... The extreme precipitation value for each cell; Indicates the first The shape parameters of the GPD distribution fitted by each cell; Indicates the first The scale parameter of the GPD distribution fitted by each cell; Indicates the first The location parameters of the GPD distribution fitted by each cell; Step 4.2: Based on the generalized Pareto distribution parameters fitted to each cell, calculate the quantile probability of extreme precipitation in the region where each cell is located: In the formula, Indicates the first Extreme precipitation values ​​for each cell The corresponding quantile probability; Step 4.3: Organize the precipitation data into... ,in This represents the number of samples input to the PA-ConvLSTM convolutional long short-term memory network. This represents the number of days in historical precipitation data; the cell space is defined as the matrix range. For the length of the cell space, For the width of the cell space, PA-ConvLSTM first encodes the spatiotemporal dependency features of historical precipitation data based on a convolutional long short-term memory network: in, Indicates that the input data is in A slice of time; and They represent and The output hidden state after passing through ConvLSTM; , , and Indicates ConvLSTM in Input gates, forget gates, output gates, and cell states at each time step; These are the learnable parameters of PA-ConvLSTM; and These represent the sigmoid and tanh activation functions, respectively. This represents the dot product operator; This represents the convolution operator; Step 4.4: Based on the output hidden state of ConvLSTM, PA-ConvLSTM applies a position-based affine transformation module to perceive the spatial differences in the hidden states of different cells, and outputs the spatially corrected hidden state. : In the formula, Represents the scale parameter of the position-based affine transformation module; This represents the offset parameter of the position-based affine transformation module. and All of these are learnable parameters in PA-ConvLSTM; Step 4.5, Hidden State Based on Spatial Correction PA-ConvLSTM uses a fully connected layer to predict the quantile probability of precipitation and GPD distribution at the next time step. In the formula, This indicates that PA-CovLSTM is in The predicted rainfall for each cell at that time; This indicates that PA-CovLSTM is in The quantile probability of the GPD distribution of each cell at time t; Step 4.6: Based on the predicted and actual values ​​of precipitation and GPD distribution quantile probabilities obtained from PA-CovLSTM in the training set TS, define the loss function. Calculation error: In the formula, This represents the number of samples in the training set TS; and Let represent the true values ​​of precipitation and GPD distribution quantile probability of the j-th sample in the training set TS, respectively. and Let represent the predicted values ​​of precipitation and GPD distribution quantile probability for the j-th sample in the training set TS, respectively. Represents the adjustable hyperparameters of the loss function; L2 norm operator; Step 4.7: Based on the backpropagation algorithm, the error calculated by the loss function is passed to each learnable parameter in PA-ConvLSTM, and the learnable parameters are updated based on the Adam optimizer. Step 4.8, iterate through steps 4.6 to 4.7 until the error value of the loss function on the validation set VS reaches the minimum value, stop the iteration process, save the learnable parameters at this time, and you will get the trained PA-ConvLSTM.

4. The method according to claim 3, characterized in that, The steps for employing different strategies for non-precipitation, ordinary precipitation, and extreme precipitation states include: When the cell state is predicted to be non-precipitation, the precipitation prediction value of that cell is set to zero. When the cell state is predicted to be normal precipitation, the precipitation prediction value directly output by PA-ConvLSTM is used. When the cell state is predicted to be an extreme precipitation state, the predicted extreme precipitation value of the cell is obtained by using the GPD quantile probability prediction value output by PA-ConvLSTM and combining it with the GPD distribution parameters pre-fitted for the cell.

5. The method according to claim 4, characterized in that, The steps for obtaining the spatial distribution of precipitation over the next few days include: The cell state and its corresponding state transition type predicted by the second cellular automaton model for the next day are used as inputs for daily iterative prediction. Different strategies are applied to the daily prediction results to calculate the spatial distribution of precipitation over the next few days.

Images