Mesoscale convection parameter optimization method and system based on genetic algorithm

Abstract

The present invention provides a mesoscale convective parameter optimization method and system based on a genetic algorithm, which relates to the field of meteorological forecasting technology. The method comprises: modeling the precipitation evolution state through a neural ordinary differential equation, identifying the dynamical system parameters in combination with variational inference, extracting features using a spatiotemporal heterogeneous graph neural network and a diffusion probability model, correcting parameter thresholds using a physics-guided neural network, and realizing parameter optimization by constructing an optimization objective function through a deep neural network. The method can improve the accuracy of precipitation forecast, reduce forecast errors, and has strong adaptability and generalization capabilities.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological forecasting, and in particular to a method and system for optimizing mesoscale convection parameters based on a genetic algorithm. Background Art

[0002] Mesoscale convection is an important process that affects the evolution of weather systems. The parameterization of mesoscale convection mainly relies on physical parameterization schemes to describe the convection process. These schemes contain a large number of parameters that need to be optimized. Traditional parameter optimization methods mainly adjust parameters based on expert experience and statistical analysis, and also use some machine learning methods to assist the parameter optimization process.

[0003] With the development of deep learning technology, researchers have begun to apply neural networks to meteorological parameter optimization. Existing deep learning methods mainly learn the mapping relationship between parameters and forecast results directly from historical data by building end-to-end prediction models. However, existing mesoscale convective parameter optimization methods still lack in-depth modeling of the dynamic characteristics of precipitation processes, have difficulty effectively handling the high dimensionality and nonlinear characteristics of the parameter space, and rarely consider physical constraints and parameter uncertainties, resulting in the optimized parameter combinations not conforming to physical laws.

[0004] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the Invention

[0005] The embodiments of the present invention provide a method and system for optimizing mesoscale convection parameters based on a genetic algorithm, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of an embodiment of the present invention provides a method for optimizing mesoscale convection parameters based on a genetic algorithm, comprising:

[0007] The historical precipitation process is modeled by the dynamic system using a pre-set neural ordinary differential equation to obtain the precipitation evolution state equation. The variational inference method is used to identify the dynamic system parameters and map them to the convection parameter space to obtain the parameter initial threshold. The multidimensional features of precipitation data and atmospheric environment data are extracted through a spatiotemporal heterogeneous graph neural network. The extracted features are denoised using a diffusion probability model to generate parameter influence features. The physical constraints of the convection process are extracted from the parameter influence features through a physics-guided neural network and the parameter initial threshold is corrected. The corrected parameter initial threshold is fused with the precipitation evolution sequence corresponding to the precipitation evolution state equation using a cross-modal visual attention network to obtain the precipitation control feature tensor and train the optimal prediction network. Based on the optimal prediction network and Monte Carlo tree search sampling, a parameter sensitivity distribution map is generated. Spectral clustering is performed to construct a parameter initialization manifold and a meta-learning-driven genetic algorithm is used to generate the first generation of candidate parameter combinations.

[0008] Based on the first generation of candidate parameter combinations, a precipitation forecast simulation is performed to generate precipitation output data, precipitation features are extracted and channel dependencies are calculated for the extracted feature graph to obtain a precipitation spatial feature graph and precipitation evolution law is extracted to obtain a time series feature sequence, the causal relationship between the time series feature sequences is modeled through a causal convolutional network, a time series causal graph is generated and structural similarity is calculated through a graph matching network to obtain a time series association matrix, the precipitation spatial feature graph, the time series association matrix and the precipitation control feature tensor are added to a feature fusion module based on a capsule network to generate a precipitation feature vector, the energy difference between the precipitation feature vector and the measured precipitation data is calculated through an energy-based contrastive learning framework to obtain feature matching, and the forecast effect is scored through a hierarchical fluctuation evaluation network, and the scoring results of different intensity intervals are adaptively fused to obtain a comprehensive forecast score;

[0009] An optimization objective function is constructed based on a deep neural network, a multi-objective fitness function based on uncertainty is constructed according to the comprehensive forecast score, and the uncertainty of the optimization objective function is modeled through a Gaussian process, a dynamic crossover operator based on a neural ordinary differential equation is constructed and a parameter evolution trajectory is constructed, the crossover position and probability are determined based on the parameter evolution trajectory, a search strategy network is constructed according to an offline reinforcement learning algorithm, the strategy network is trained according to the historical parameter optimization trajectory, and a crossover operation is performed on the parameter combination whose fitness value is higher than a preset fitness threshold to generate a second-generation parameter combination, a parameter search guided by the strategy network is performed on the second-generation parameter combination to obtain an optimized parameter combination, the stability of the optimized parameter combination is calculated in combination with a variational inference algorithm, and the optimized parameter combination with the highest stability is selected as the optimal parameter combination output.

[0010] In an optional embodiment,

[0011] The historical precipitation process is modeled by the dynamic system using a pre-set neural ordinary differential equation to obtain the precipitation evolution state equation. The variational inference method is used to identify the dynamic system parameters and map them to the convection parameter space to obtain the parameter initial threshold. The multidimensional features of precipitation data and atmospheric environment data are extracted through a spatiotemporal heterogeneous graph neural network. The extracted features are denoised using a diffusion probability model to generate parameter influence features. The physical constraints of the convection process are extracted from the parameter influence features through a physics-guided neural network and the parameter initial threshold is corrected. The modified parameter initial threshold is fused with the precipitation evolution sequence corresponding to the precipitation evolution state equation using a cross-modal visual attention network to obtain the precipitation control feature tensor and train the optimal prediction network. Based on the optimal prediction network and Monte Carlo tree search sampling, a parameter sensitivity distribution map is generated. Spectral clustering is performed to construct a parameter initialization manifold and a meta-learning-driven genetic algorithm is used to generate the first generation of candidate parameter combinations, including:

[0012] A dynamic system model is performed on the historical precipitation process using a preset Neural Ordinary Differential Equation. The historical precipitation process includes precipitation intensity, precipitation duration, and precipitation spatial distribution data. The historical precipitation process data is input into the Neural Ordinary Differential Equation in time series. An adaptive step-size solver is used to perform time series integration on the historical precipitation process data and fit the dynamic evolution trend. The precipitation evolution state equation that describes the temporal evolution characteristics of precipitation intensity is output;

[0013] Inputting the precipitation evolution state equation into a variational inference method for parameter extraction, encoding and mapping the precipitation evolution state equation into a parameter distribution space to obtain dynamic system parameters, decoding and reconstructing the dynamic system parameters into a state equation, constructing an optimization target by calculating the error between the reconstructed state equation and the original state equation, and the difference between the parameter distribution and the prior distribution, iterating the optimization until the error converges to extract the dynamic system parameters, and converting the dynamic system parameters into a convection parameter space through physical mapping to obtain an initial parameter threshold;

[0014] Collect precipitation data and atmospheric environment data, including temperature, relative humidity, and wind field elements. Divide the forecast area into grid points according to a preset spatial resolution. Construct a graph structure based on the grid points. The nodes of the graph structure store the precipitation data and atmospheric environment data corresponding to the grid point positions. The edges of the graph structure calculate the spatial correlation strength based on the spatial distance between the grid points and the correlation with meteorological elements. Input the graph structure into a preset spatiotemporal heterogeneous graph neural network. Node and edge features are extracted through graph convolution operations and a message passing mechanism. The time dimension information is then integrated to obtain multidimensional features.

[0015] Denoising the multidimensional features using a diffusion probability model, gradually adding Gaussian noise to the features according to a preset step size during a forward diffusion process, wherein the noise variance increases exponentially with the number of diffusion steps, inputting the noisy feature sequence into a pre-set denoising network, and gradually recovering the feature information through a reverse diffusion process, wherein the denoising network retains multi-scale feature details during the reverse diffusion process, and ultimately generating parameter influence features reflecting the mechanism of the parameter's influence on precipitation;

[0016] Inputting the parameter influence characteristics into a pre-set physics-guided neural network to extract physical constraints of the convection process, wherein the physical constraints include convection triggering conditions, convection development conditions, and parameter validity conditions; and evaluating the physical rationality of the initial parameter thresholds based on the physical constraints of the convection process and performing corrections;

[0017] The modified parameter initial threshold and the precipitation evolution sequence corresponding to the precipitation evolution state equation are input into a pre-set cross-modal visual attention network, the internal correlation of the parameters is calculated through the self-attention mechanism, and the correlation between the parameters and the precipitation evolution sequence is calculated through the cross-attention mechanism. The parameter information and the precipitation evolution information are projected into a unified feature space and fused with the attention weight to generate a precipitation control feature tensor containing the parameter control effect and the precipitation evolution characteristics;

[0018] An optimal prediction network is obtained based on the precipitation control feature tensor training, the precipitation control feature tensor is input into the optimal prediction network to extract feature dependencies, and feature sampling is performed in combination with the Monte Carlo tree search method. The feature sampling explores the parameter space through tree node expansion, calculates the node value through backtracking update, and generates a parameter sensitivity distribution map describing the degree of parameter influence based on the sampling results;

[0019] Spectral clustering is performed on the parameter sensitivity distribution map, an affinity matrix is constructed based on the map data, the eigenvalues and eigenvectors of the affinity matrix are calculated, the parameter sensitivity distribution is projected into the characteristic subspace for clustering analysis, and a parameter initialization manifold describing the geometric structure of the parameter distribution is constructed based on the clustering results; within the parameter initialization manifold, a meta-learning-driven genetic algorithm is used to learn historical optimization experience and quickly adapt to the current parameter search task, and the first generation of candidate parameter combinations that meet the physical meaning are generated in combination with the parameter constraints.

[0020] In an optional embodiment,

[0021] An optimal prediction network is obtained based on the precipitation control feature tensor training, the precipitation control feature tensor is input into the optimal prediction network to extract feature dependencies, and feature sampling is performed in combination with the Monte Carlo tree search method. The feature sampling explores the parameter space through tree node expansion, calculates the node value through backtracking update, and generates a parameter sensitivity distribution map describing the degree of parameter influence based on the sampling results, including:

[0022] The precipitation control feature tensor is divided into a training set and a validation set according to a time window, the training set is input into a preset optimal prediction network for training, an encoder of the optimal prediction network is controlled to calculate the similarity between feature vectors through a self-attention mechanism to obtain attention weights, features are weightedly aggregated based on the attention weights to obtain feature dependencies, and a decoder of the optimal prediction network is controlled to construct a mapping relationship between features and prediction targets through a cross-attention mechanism;

[0023] An initial learning rate is set to 0.001 and the optimal prediction network is optimized. When the verification loss on the verification set does not decrease for five consecutive iterations, the learning rate is adjusted to 0.1 times the original value, the gradient norm is limited to a preset gradient norm threshold, and gradient clipping is performed. When the prediction error on the verification set is monitored to be less than a preset error threshold, the network training is stopped to obtain the current parameter state.

[0024] Constructing a Monte Carlo search tree and initializing a root node, writing the current parameter state into the root node, inputting the parameter state into the trained optimal prediction network to perform evaluation, expanding and exploring the parameter space through tree nodes, screening multiple parameter adjustment solutions with the highest scores based on the evaluation results to construct child nodes, and writing new parameter values and state scores into the child nodes;

[0025] Starting from the leaf node of the search tree, perform a backtracking update of the node score along the search path upward, update the score of the current child node according to the score and the number of visits of the child node, and assign exploration rewards to the child node according to the number of visits;

[0026] Based on the sampling results, a statistical analysis is performed on the number of visits and score distributions of sub-nodes corresponding to different parameter values in the search tree, the degree of exploration of the parameter values is determined according to the number of visits, and the importance of the parameter values is determined according to the score distribution. The number of visits and the score distribution are fused to calculate a quantitative index of the parameter influence degree, and the quantitative index is used to generate a parameter sensitivity distribution map describing the parameter influence degree.

[0027] In an optional embodiment,

[0028] Based on the first generation of candidate parameter combinations, precipitation forecast simulation is performed to generate precipitation output data, precipitation features are extracted and channel dependencies are calculated for the extracted feature graphs to obtain precipitation spatial feature graphs and precipitation evolution laws are extracted to obtain time series feature sequences. The causal relationship between the time series feature sequences is modeled through a causal convolutional network, a time series causal graph is generated and structural similarity is calculated through a graph matching network to obtain a time series association matrix. The precipitation spatial feature graph, the time series association matrix and the precipitation control feature tensor are added to a feature fusion module based on a capsule network to generate a precipitation feature vector. The energy difference between the precipitation feature vector and the measured precipitation data is calculated through an energy-based contrastive learning framework to obtain feature matching and score the forecast effect through a hierarchical fluctuation evaluation network. The scoring results of different intensity intervals are adaptively fused to obtain a comprehensive forecast score including:

[0029] constructing a test parameter combination input matrix based on the first generation candidate parameter combinations, performing normalization processing on the test parameter combination input matrix to obtain a normalized parameter matrix, and inputting the normalized parameter matrix into a preset numerical forecast model to perform precipitation forecast simulation and generate precipitation output data;

[0030] Constructing multiple parallel feature extraction branches to extract precipitation features from the precipitation output data, wherein the feature extraction branches extract precipitation features of different scales through a multi-layer convolutional network, perform batch normalization and activation function processing after each convolution layer, and reduce the feature dimension through maximum pooling to obtain a feature map;

[0031] Constructing a channel attention module to calculate the channel dependency of the feature map, compressing the feature map through global average pooling to obtain a channel descriptor, processing the channel descriptor using a multi-layer fully connected network to obtain a channel weight, and reweighting the feature map based on the channel weight to obtain a precipitation spatial feature map;

[0032] The precipitation field features of multiple forecast times are connected in series to construct a time series feature sequence. The precipitation evolution law is extracted through a multi-layer dilated causal convolutional network. The dilated causal convolutional network expands the receptive field by increasing the dilation rate layer by layer and adds residual connections to prevent gradient vanishing. The time series features are constructed into a time series causal graph. The causal relationship strength between nodes in the time series causal graph is calculated based on the attention mechanism. The measured precipitation data is constructed into a reference causal graph. The structural similarity between the time series causal graph and the reference causal graph is calculated using a graph matching network to obtain a time series correlation matrix.

[0033] Constructing a feature fusion module based on a capsule network, the feature fusion module includes multiple primary capsule layers and advanced capsule layers. The primary capsule layers process the precipitation spatial feature map, the temporal correlation matrix, and the precipitation control feature tensor respectively, and integrate multi-source features in the advanced capsule layer through a dynamic routing algorithm to obtain a precipitation feature vector;

[0034] An energy-based contrastive learning framework is constructed. The precipitation feature vector and the measured precipitation data are input into a feature encoder and mapped to a contrastive learning space. Positive and negative sample pairs are constructed in preset batches. The feature matching degree is calculated based on the energy difference between the positive and negative sample pairs. A hierarchical fluctuation assessment network is constructed. The hierarchical fluctuation assessment network includes multiple evaluation branches for different precipitation intensity intervals. Each evaluation branch extracts the forecast features of the corresponding intensity interval through a multi-layer convolutional network and an attention mechanism to obtain a branch score. The feature matching degree is used as a weight adjustment factor for each branch score to fuse the branch scores to obtain a comprehensive forecast score.

[0035] In an optional embodiment,

[0036] Construct a feature fusion module based on a capsule network. The feature fusion module includes multiple primary capsule layers and advanced capsule layers. The primary capsule layers process the precipitation spatial feature map, the temporal correlation matrix, and the precipitation control feature tensor respectively. The dynamic routing algorithm is used to integrate multi-source features in the advanced capsule layer to obtain a precipitation feature vector. The method includes:

[0037] Constructing a feature fusion module based on a capsule network, the feature fusion module includes multiple primary capsule layers and multiple high-level capsule layers. The primary capsule layers respectively process a precipitation spatial feature map, a temporal association matrix, and a precipitation control feature tensor. The primary capsule layers include multiple capsule units with fixed dimensions. Feature channels of the precipitation spatial feature map are reorganized into multiple groups and mapped to a capsule space through a feature transformation matrix to obtain a first group of capsule representations. The temporal association matrix is divided into multiple overlapping sub-matrices and mapped to a capsule space through a feature transformation matrix to obtain a second group of capsule representations. The precipitation control feature tensor is grouped along a physical quantity dimension and mapped to a capsule space through a feature transformation matrix to obtain a third group of capsule representations.

[0038] The advanced capsule layer includes a plurality of capsule units of fixed dimensions, and a coupling coefficient matrix is initialized to establish a connection relationship between the primary capsule layer and the advanced capsule layer. The dimension of the coupling coefficient matrix is the number of primary capsules multiplied by the number of advanced capsules;

[0039] Mapping the capsule representation of the primary capsule layer into a prediction vector through a feature transformation matrix, calculating a temporary representation of the high-level capsule layer based on the prediction vector and the coupling coefficient matrix, and performing nonlinear compression on the temporary representation to obtain an activation representation of the high-level capsule layer;

[0040] The similarity of the predicted vector is calculated based on the activation representation of the high-level capsule layer, the coupling coefficient in the coupling coefficient matrix is updated according to the similarity and normalization is performed, the connection relationship between the primary capsule layer and the high-level capsule layer is repeatedly optimized through multiple rounds of iterations using a dynamic routing algorithm, and the temporary representation and activation representation of the high-level capsule layer are sequentially connected based on the connection relationship to generate a precipitation feature vector, wherein the precipitation feature vector includes precipitation spatial distribution information, temporal evolution information, and physical control information.

[0041] In an optional embodiment,

[0042] An optimization objective function is constructed based on a deep neural network, a multi-objective fitness function based on uncertainty is constructed according to the comprehensive forecast score, and the uncertainty of the optimization objective function is modeled through a Gaussian process, a dynamic crossover operator based on a neural ordinary differential equation is constructed and a parameter evolution trajectory is constructed, a crossover position and probability are determined based on the parameter evolution trajectory, a search strategy network is constructed according to an offline reinforcement learning algorithm, the strategy network is trained according to a historical parameter optimization trajectory, and a crossover operation is performed on a parameter combination whose fitness value is higher than a preset fitness threshold to generate a second-generation parameter combination, a parameter search guided by the strategy network is performed on the second-generation parameter combination to obtain an optimized parameter combination, the stability of the optimized parameter combination is calculated in combination with a variational inference algorithm, and the optimized parameter combination with the highest stability is selected as the optimal parameter combination output, which includes:

[0043] Constructing an optimization objective function based on a deep neural network, wherein the deep neural network consists of an input layer, a hidden layer, and an output layer, inputting the comprehensive forecast score into the input layer, mapping the outputs of the input layer neurons to the hidden layer via a weight matrix, performing a rectified linear activation operation on the outputs of the hidden layer to introduce nonlinear characteristics, and mapping the activated features to the output layer via a fully connected layer to generate an objective function value in scalar form;

[0044] Constructing an uncertainty-based multi-objective fitness function based on the comprehensive forecast score, calculating the similarity between sample points using a radial basis kernel function to establish a Gaussian process covariance matrix, iteratively optimizing the signal variance and length scale hyperparameters of the Gaussian process using a maximum likelihood estimation method, calculating the fitness expectation value and uncertainty index of the parameter configuration to be evaluated using the Gaussian process model, and weightedly combining the fitness expectation value and the uncertainty index to generate a multi-objective fitness function value;

[0045] A dynamic crossover operator based on a neural network of ordinary differential equations is constructed, and a multilayer perceptron network is designed as a parameter change rate calculation unit. The multilayer perceptron network outputs the rate of change of the parameter over time by subjecting the current state of the parameter to multi-layer nonlinear transformations. The Runge-Kutta algorithm is used to numerically solve the neural network of ordinary differential equations to obtain a parameter evolution trajectory. The curvature value of the parameter evolution trajectory at each time point is calculated based on the time series data of the parameter evolution trajectory. The time point where the curvature value is greater than a preset curvature threshold is determined as the crossover position. The corresponding crossover probability is set according to the curvature value, wherein the position with a larger curvature value is assigned a higher crossover probability.

[0046] A search strategy network is constructed based on an offline reinforcement learning algorithm. A dual network structure consisting of a strategy network and a value network is designed. The strategy network uses a multi-layer convolutional structure to extract parameter state features and output crossover probability and mutation strength. The value network uses a fully connected structure to evaluate operation value. State transition sequences and reward signals from historical parameter optimization trajectories are collected to construct an experience replay pool. Training data is batch-sampled from the experience replay pool to optimize the network parameters of the dual network structure.

[0047] Inputting a parameter combination whose fitness value is higher than a preset fitness threshold into the policy network, obtaining a crossover probability and a mutation strength, performing a simulated binary crossover operation at a high curvature position of the parameter evolution trajectory to generate a second-generation parameter combination, performing a parameter search guided by the policy network on the second-generation parameter combination, inputting the parameter combination into the policy network to obtain a search direction and a step size, updating the parameters along the search direction and calculating the objective function value, determining whether to accept the parameter update based on a change in the objective function value and an acceptance probability output by the policy network, and performing multiple rounds of iterative search on each set of parameters to obtain an optimized parameter combination;

[0048] A variational distribution in the form of a diagonal Gaussian is constructed in combination with the variational inference algorithm. The optimized parameter combination is input into the variational distribution to calculate the posterior probability. The variational parameters are iteratively optimized using the stochastic gradient descent algorithm to minimize the relative entropy between the variational distribution and the true posterior distribution. The variance of the variational distribution is calculated as a measure of parameter stability, and the optimized parameter combination with the smallest variance is selected as the optimal parameter combination output.

[0049] In an optional embodiment,

[0050] A dynamic crossover operator based on the Neural Ordinary Differential Equation is constructed, and a multilayer perceptron network is designed as a parameter change rate calculation unit. The multilayer perceptron network outputs the parameter change rate over time by subjecting the current state of the parameter to multi-layer nonlinear transformations. The Runge-Kutta algorithm is used to numerically solve the Neural Ordinary Differential Equation to obtain the parameter evolution trajectory, including:

[0051] A multilayer perceptron network is designed as a parameter change rate calculation unit. The multilayer perceptron network consists of an input layer, multiple hidden layers, and an output layer. The input layer receives the current state of the parameter. The number of neurons in the multiple hidden layers decreases layer by layer, and a normalization layer is set at the output end of each layer. The normalized features are nonlinearly transformed using a hyperbolic tangent function. The output layer maps the transformed features through a fully connected layer to obtain the rate of change of the parameter over time.

[0052] Extracting a parameter evolution sequence from a historical optimization database, normalizing parameter values in the parameter evolution sequence and calculating parameter changes between adjacent sampling points, using the normalized parameter values as inputs to the multilayer perceptron network, and training the multilayer perceptron network using the parameter changes as supervisory signals, and stopping training when the validation set loss does not decrease for multiple consecutive rounds;

[0053] The trained multilayer perceptron network is embedded in an ordinary differential equation framework to construct a core component of a dynamic crossover operator, wherein the dynamic crossover operator includes a parameter evolution part and a crossover operation part. In the parameter evolution part, the parameter configuration to be crossed is used as the initial state, and the right-hand side of the neural ordinary differential equation is constructed by the rate of change output by the multilayer perceptron network. The neural ordinary differential equation is numerically solved, and the integration interval is divided into an initial step sequence. The parameter state is input into the multilayer perceptron network at different time points to obtain a change rate estimate. The intermediate point prediction and weighted combination of the Runge-Kutta algorithm are used to calculate the parameter state at the next time point. The integration step is adaptively adjusted according to the truncation error between adjacent time points. The neural ordinary differential equation is numerically solved to obtain a candidate parameter evolution trajectory.

[0054] The candidate parameter evolution trajectories are organized in time series, the average rate of change, rate of change variance, and state difference of the candidate parameter evolution trajectories are calculated, and key turning points on the candidate parameter evolution trajectories are identified. Based on the key turning points, the convergence characteristics of the candidate parameter evolution trajectories are evaluated according to the attenuation trend of the parameter change rate, the wandering characteristics in the state space, and the stability of the trajectory endpoint. The unconverged trajectories are re-solved by adjusting the integration interval, and the candidate parameter evolution trajectories are projected into the physical constraint space. The physical rationality of the candidate parameter evolution trajectories is verified through parameter value range inspection, physical correlation evaluation, and simplified physical model.

[0055] The candidate parameter evolution trajectories that pass the convergence test and the physical rationality test are scored for quality, and the parameter evolution trajectories are screened according to the trajectory smoothness, stability and physical interpretability and stored in the optimization experience database.

[0056] A second aspect of an embodiment of the present invention provides a mesoscale convection parameter optimization system based on a genetic algorithm, comprising:

[0057] The first unit is used to perform dynamic system modeling of the historical precipitation process through a preset neural ordinary differential equation to obtain a precipitation evolution state equation, identify dynamic system parameters in combination with a variational inference method and map them to the convection parameter space to obtain parameter initial thresholds, extract multidimensional features of precipitation data and atmospheric environment data through a spatiotemporal heterogeneous graph neural network, perform denoising on the extracted features to generate parameter influence features through a diffusion probability model, extract physical constraints of the convection process from the parameter influence features through a physics-guided neural network and correct the parameter initial thresholds, combine the cross-modal visual attention network to fuse the corrected parameter initial thresholds and the precipitation evolution sequence corresponding to the precipitation evolution state equation to obtain a precipitation control feature tensor and train an optimal prediction network, generate a parameter sensitivity distribution map based on the optimal prediction network and Monte Carlo tree search sampling, perform spectral clustering to construct a parameter initialization manifold, and generate a first-generation candidate parameter combination in combination with a meta-learning-driven genetic algorithm;

[0058] The second unit is used to perform precipitation forecast simulation based on the first-generation candidate parameter combination to generate precipitation output data, extract precipitation features and calculate channel dependencies on the extracted feature graph to obtain a precipitation spatial feature graph and extract precipitation evolution laws to obtain a time series feature sequence, model the causal relationship between the time series feature sequences through a causal convolutional network, generate a time series causal graph and calculate structural similarity through a graph matching network to obtain a time series association matrix, add the precipitation spatial feature graph, the time series association matrix and the precipitation control feature tensor to a feature fusion module based on a capsule network to generate a precipitation feature vector, calculate the energy difference between the precipitation feature vector and the measured precipitation data through an energy-based contrastive learning framework, obtain feature matching, and score the forecast effect through a hierarchical fluctuation evaluation network, and adaptively fuse the scoring results of different intensity intervals to obtain a comprehensive forecast score;

[0059] The third unit is used to construct an optimization objective function based on a deep neural network, construct a multi-objective fitness function based on uncertainty according to the comprehensive forecast score and model the uncertainty of the optimization objective function through a Gaussian process, construct a dynamic crossover operator based on a neural ordinary differential equation and construct a parameter evolution trajectory, determine the crossover position and probability based on the parameter evolution trajectory, construct a search strategy network according to an offline reinforcement learning algorithm, train the strategy network according to the historical parameter optimization trajectory and perform a crossover operation on the parameter combination whose fitness value is higher than a preset fitness threshold to generate a second-generation parameter combination, perform a parameter search guided by the strategy network on the second-generation parameter combination to obtain an optimized parameter combination, calculate the stability of the optimized parameter combination in combination with a variational inference algorithm, and select the optimized parameter combination with the highest stability as the optimal parameter combination output.

[0060] According to a third aspect of the embodiments of the present invention,

[0061] An electronic device is provided, comprising:

[0062] processor;

[0063] a memory for storing processor-executable instructions;

[0064] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0065] According to a fourth aspect of the embodiments of the present invention,

[0066] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0067] In the present invention, the dynamic system modeling of the historical precipitation process is carried out by using the neural ordinary differential equation, and the variational inference method is combined to identify the dynamic system parameters. At the same time, the spatiotemporal heterogeneous graph neural network and the diffusion probability model are used to extract features, thereby realizing deep mining of precipitation data and atmospheric environment data, improving the initialization accuracy of convection parameters, and providing a better search starting point for subsequent optimization. The capsule network is used to realize multidimensional feature fusion of precipitation spatial feature map, time series correlation matrix and precipitation control feature tensor, and the feature matching degree is evaluated by the energy-based contrast learning framework. The hierarchical fluctuation assessment network is combined to perform interval scoring, thereby realizing a refined evaluation of the forecast effect of different precipitation intensity intervals, improving the rationality and reliability of the forecast scoring, constructing an optimization objective function based on a deep neural network, modeling uncertainty through a Gaussian process, constructing a dynamic crossover operator in combination with the neural ordinary differential equation, and guiding parameter search by using an offline reinforcement learning algorithm. The parameter stability is evaluated by the variational inference algorithm, thereby realizing intelligent guidance and stability guarantee of the parameter optimization process, and improving the reliability and applicability of the optimization results. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 Schematic diagram of the flow of a method for optimizing mesoscale convection parameters based on a genetic algorithm according to an embodiment of the present invention;

[0069] Figure 2 Schematic diagram of the structure of a mesoscale convection parameter optimization system based on a genetic algorithm according to an embodiment of the present invention. DETAILED DESCRIPTION

[0070] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0071] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0072] Figure 1 FIG. 1 is a flow chart of a method for optimizing mesoscale convection parameters based on a genetic algorithm according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0073] S1. Use a pre-set neural ordinary differential equation to perform dynamic system modeling on the historical precipitation process to obtain the precipitation evolution state equation. Combine the variational inference method to identify the dynamic system parameters and map them to the convection parameter space to obtain the parameter initial threshold. Use a spatiotemporal heterogeneous graph neural network to extract the multidimensional features of precipitation data and atmospheric environment data. Use a diffusion probability model to denoise the extracted features to generate parameter influence features. Use a physics-guided neural network to extract the physical constraints of the convection process from the parameter influence features and correct the parameter initial threshold. Combine the cross-modal visual attention network to fuse the corrected parameter initial threshold and the precipitation evolution sequence corresponding to the precipitation evolution state equation to obtain the precipitation control feature tensor and train the optimal prediction network. Based on the optimal prediction network and Monte Carlo tree search sampling, generate a parameter sensitivity distribution map. Perform spectral clustering to construct a parameter initialization manifold and combine it with a meta-learning-driven genetic algorithm to generate the first generation of candidate parameter combinations.

[0074] The variational inference method is a technique for inferring latent variables in complex probabilistic models by optimizing the approximate posterior distribution and is widely used in Bayesian statistics and deep learning. The convection parameter space is a space used to describe and optimize the dynamic parameters of convection systems and is commonly used in fluid mechanics and meteorological modeling. The spatiotemporal heterogeneous graph neural network is a graph neural network that simultaneously models spatial structure and temporal evolution and can handle complex dependencies between nodes and edges of heterogeneous data. The diffusion probability model is a model that generates new data based on the diffusion process and generates samples by reversely simulating the diffusion path. It is used for tasks such as image generation. The physics-guided neural network is a method that incorporates physical constraints or models into the neural network structure to enhance the network's modeling ability of physical phenomena. The cross-modal visual attention network is a neural network that extracts relevant features across different modalities through an attention mechanism and is used for multimodal tasks. The Monte Carlo tree search sampling is a method that combines random sampling with a tree search strategy and is commonly used in reinforcement learning and decision-making problems to explore optimal solutions in high-dimensional spaces. The parameter sensitivity distribution map is a visualization tool that represents the sensitivity of model parameters to outputs and is used for model tuning and interpretation.

[0075] In an optional embodiment,

[0076] The historical precipitation process is modeled by the dynamic system using a pre-set neural ordinary differential equation to obtain the precipitation evolution state equation. The variational inference method is used to identify the dynamic system parameters and map them to the convection parameter space to obtain the parameter initial threshold. The multidimensional features of precipitation data and atmospheric environment data are extracted through a spatiotemporal heterogeneous graph neural network. The extracted features are denoised using a diffusion probability model to generate parameter influence features. The physical constraints of the convection process are extracted from the parameter influence features through a physics-guided neural network and the parameter initial threshold is corrected. The modified parameter initial threshold is fused with the precipitation evolution sequence corresponding to the precipitation evolution state equation using a cross-modal visual attention network to obtain the precipitation control feature tensor and train the optimal prediction network. Based on the optimal prediction network and Monte Carlo tree search sampling, a parameter sensitivity distribution map is generated. Spectral clustering is performed to construct a parameter initialization manifold and a meta-learning-driven genetic algorithm is used to generate the first generation of candidate parameter combinations, including:

[0077] A dynamic system model is performed on the historical precipitation process using a preset Neural Ordinary Differential Equation. The historical precipitation process includes precipitation intensity, precipitation duration, and precipitation spatial distribution data. The historical precipitation process data is input into the Neural Ordinary Differential Equation in time series. An adaptive step-size solver is used to perform time series integration on the historical precipitation process data and fit the dynamic evolution trend. The precipitation evolution state equation that describes the temporal evolution characteristics of precipitation intensity is output;

[0078] Inputting the precipitation evolution state equation into a variational inference method for parameter extraction, encoding and mapping the precipitation evolution state equation into a parameter distribution space to obtain dynamic system parameters, decoding and reconstructing the dynamic system parameters into a state equation, constructing an optimization target by calculating the error between the reconstructed state equation and the original state equation, and the difference between the parameter distribution and the prior distribution, iterating the optimization until the error converges to extract the dynamic system parameters, and converting the dynamic system parameters into a convection parameter space through physical mapping to obtain an initial parameter threshold;

[0079] Collect precipitation data and atmospheric environment data, including temperature, relative humidity, and wind field elements. Divide the forecast area into grid points according to a preset spatial resolution. Construct a graph structure based on the grid points. The nodes of the graph structure store the precipitation data and atmospheric environment data corresponding to the grid point positions. The edges of the graph structure calculate the spatial correlation strength based on the spatial distance between the grid points and the correlation with meteorological elements. Input the graph structure into a preset spatiotemporal heterogeneous graph neural network. Node and edge features are extracted through graph convolution operations and a message passing mechanism. The time dimension information is then integrated to obtain multidimensional features.

[0080] Denoising the multidimensional features using a diffusion probability model, gradually adding Gaussian noise to the features according to a preset step size during a forward diffusion process, wherein the noise variance increases exponentially with the number of diffusion steps, inputting the noisy feature sequence into a pre-set denoising network, and gradually recovering the feature information through a reverse diffusion process, wherein the denoising network retains multi-scale feature details during the reverse diffusion process, and ultimately generating parameter influence features reflecting the mechanism of the parameter's influence on precipitation;

[0081] Inputting the parameter influence characteristics into a pre-set physics-guided neural network to extract physical constraints of the convection process, wherein the physical constraints include convection triggering conditions, convection development conditions, and parameter validity conditions; and evaluating the physical rationality of the initial parameter thresholds based on the physical constraints of the convection process and performing corrections;

[0082] The modified parameter initial threshold and the precipitation evolution sequence corresponding to the precipitation evolution state equation are input into a pre-set cross-modal visual attention network, the internal correlation of the parameters is calculated through the self-attention mechanism, and the correlation between the parameters and the precipitation evolution sequence is calculated through the cross-attention mechanism. The parameter information and the precipitation evolution information are projected into a unified feature space and fused with the attention weight to generate a precipitation control feature tensor containing the parameter control effect and the precipitation evolution characteristics;

[0083] An optimal prediction network is obtained based on the precipitation control feature tensor training, the precipitation control feature tensor is input into the optimal prediction network to extract feature dependencies, and feature sampling is performed in combination with the Monte Carlo tree search method. The feature sampling explores the parameter space through tree node expansion, calculates the node value through backtracking update, and generates a parameter sensitivity distribution map describing the degree of parameter influence based on the sampling results;

[0084] Spectral clustering is performed on the parameter sensitivity distribution map, an affinity matrix is constructed based on the map data, the eigenvalues and eigenvectors of the affinity matrix are calculated, the parameter sensitivity distribution is projected into the characteristic subspace for clustering analysis, and a parameter initialization manifold describing the geometric structure of the parameter distribution is constructed based on the clustering results; within the parameter initialization manifold, a meta-learning-driven genetic algorithm is used to learn historical optimization experience and quickly adapt to the current parameter search task, and the first generation of candidate parameter combinations that meet the physical meaning are generated in combination with the parameter constraints.

[0085] The neural ordinary differential equation is a type of model that uses neural networks to parameterize ordinary differential equations, which can be used to model continuous-time dynamic systems. The spatial resolution is an indicator that describes the detail recognition ability of data or models in the spatial dimension, and is used to evaluate the accuracy of image or geographic data. The inverse diffusion process is a method of restoring data distribution by inversely simulating the diffusion process, which is used to generate models or denoising tasks. The feature subspace is a low-dimensional space used to represent the core information of high-dimensional feature data, helping to reduce dimensions and accelerate calculations. The meta-learning-driven genetic algorithm is an algorithm that combines meta-learning ideas to optimize genetic algorithm parameters and operations, and is used to improve the adaptability and efficiency of the genetic algorithm.

[0086] Dynamical system modeling is performed based on historical precipitation data. Historical data, including precipitation intensity, duration, and spatial distribution, are processed using a neural ordinary differential equation (ODE). Specifically, a Dormand-Prince adaptive step-size solver is used to perform time series integration. Using rainfall data from the 2020 flood season in a specific region as an example, hourly rainfall intensity data is input into the DE as a time series. The precipitation evolution equation of state is obtained through integration over 100 time steps.

[0087] The dynamical system parameters were extracted using a variational inference method. The state equations were mapped to a latent space using an encoder, yielding dynamical system parameters that follow a Gaussian distribution. The reconstruction error and KL divergence were used as optimization targets, and the optimization process was iterated 5000 times until convergence. The extracted dynamical system parameters were converted into initial thresholds in the convective parameter space through physical mapping, including key parameters such as the convective trigger temperature and relative humidity threshold.

[0088] A spatiotemporal heterogeneous graph network is constructed to extract features. The forecast area is divided into grid points at a spatial resolution of 10 kilometers, and a graph structure is constructed based on these grid points. Nodes store data such as precipitation, temperature, relative humidity, wind speed and direction at that location. Edge weights are determined by spatial distance and meteorological factor correlation coefficients. Multidimensional features are extracted using three graph convolutional layers and two temporal convolutional layers.

[0089] Based on feature extraction, a diffusion probability model was used to generate parameter influence features. A diffusion step size of 1000 was set, and the noise variance was exponentially increased starting from 0.01. The denoising network used a U-Net architecture consisting of four downsampling blocks and four upsampling blocks, using skip connections to preserve multi-scale features.

[0090] A physics-guided neural network is used to extract convection process constraints. The network contains three physical constraint modules, which verify the physical rationality of parameters such as convection trigger temperature, relative humidity, and vertical velocity, and correct parameter values that do not meet the constraints.

[0091] Parameters and precipitation information are fused via a cross-modal visual attention network. The attention network consists of six self-attention layers and three cross-attention layers, with eight attention heads. Parameter features and the precipitation evolution sequence are projected into a 512-dimensional feature space for fusion, generating a precipitation control feature tensor.

[0092] An optimal prediction network consisting of three fully connected layers was trained based on the feature tensor. A Monte Carlo tree search (MTS) sampling algorithm was used to generate parameter sensitivity distributions, with a search tree depth of 10 and 1000 sampling cycles. Spectral clustering was performed on the sensitivity distributions with a cluster count of 5 to construct a parameter initialization manifold. Finally, a meta-learning-driven genetic algorithm was used with a population size of 100 and an evolutionary number of 50 to generate candidate parameter combinations that satisfied the physical constraints.

[0093] In this embodiment, by combining the neural ordinary differential equations with variational inference, the dynamic evolution characteristics of the precipitation process can be accurately captured, and the system parameters with physical significance can be extracted, providing a reliable basis for subsequent parameter optimization. The combination of spatiotemporal heterogeneous graph networks and diffusion probability models can effectively extract the multi-scale characteristics of precipitation data and atmospheric environment data, and obtain feature representations reflecting the parameter influence mechanism through denoising, thereby improving the accuracy and robustness of feature extraction. Based on the physics-guided neural network and cross-modal attention mechanism, the accurate characterization of parameter physical constraints and the deep integration of parameter precipitation information are achieved. Combined with the Monte Carlo tree search and spectral clustering methods, the parameter initialization manifold is reliably constructed, providing an efficient search space for parameter optimization.

[0094] In an optional embodiment,

[0095] An optimal prediction network is obtained based on the precipitation control feature tensor training, the precipitation control feature tensor is input into the optimal prediction network to extract feature dependencies, and feature sampling is performed in combination with the Monte Carlo tree search method. The feature sampling explores the parameter space through tree node expansion, calculates the node value through backtracking update, and generates a parameter sensitivity distribution map describing the degree of parameter influence based on the sampling results, including:

[0096] The precipitation control feature tensor is divided into a training set and a validation set according to a time window, the training set is input into a preset optimal prediction network for training, an encoder of the optimal prediction network is controlled to calculate the similarity between feature vectors through a self-attention mechanism to obtain attention weights, features are weightedly aggregated based on the attention weights to obtain feature dependencies, and a decoder of the optimal prediction network is controlled to construct a mapping relationship between features and prediction targets through a cross-attention mechanism;

[0097] An initial learning rate is set to 0.001 and the optimal prediction network is optimized. When the verification loss on the verification set does not decrease for five consecutive iterations, the learning rate is adjusted to 0.1 times the original value, the gradient norm is limited to a preset gradient norm threshold, and gradient clipping is performed. When the prediction error on the verification set is monitored to be less than a preset error threshold, the network training is stopped to obtain the current parameter state.

[0098] Constructing a Monte Carlo search tree and initializing a root node, writing the current parameter state into the root node, inputting the parameter state into the trained optimal prediction network to perform evaluation, expanding and exploring the parameter space through tree nodes, screening multiple parameter adjustment solutions with the highest scores based on the evaluation results to construct child nodes, and writing new parameter values and state scores into the child nodes;

[0099] Starting from the leaf node of the search tree, perform a backtracking update of the node score along the search path upward, update the score of the current child node according to the score and the number of visits of the child node, and assign exploration rewards to the child node according to the number of visits;

[0100] Based on the sampling results, a statistical analysis is performed on the number of visits and score distributions of sub-nodes corresponding to different parameter values in the search tree, the degree of exploration of the parameter values is determined according to the number of visits, and the importance of the parameter values is determined according to the score distribution. The number of visits and the score distribution are fused to calculate a quantitative index of the parameter influence degree, and the quantitative index is used to generate a parameter sensitivity distribution map describing the parameter influence degree.

[0101] The optimal prediction network is a neural network structure that learns the optimal strategy of historical data to make future predictions, and is used for time series prediction and control optimization. The feature dependency refers to the correlation or causal relationship between data features, which is used to enhance feature selection and modeling effects. The gradient norm is the size of the gradient in neural network training, which is used to measure gradient changes and optimization direction. The gradient clipping is a technology that controls the range of gradient values to prevent gradient explosion or vanishing problems and ensure training stability. The exploration reward is a reward strategy used in reinforcement learning to encourage exploration of unvisited states or actions, thereby improving learning efficiency and comprehensiveness.

[0102] Data preprocessing was performed on the precipitation control feature tensor. The raw data, including meteorological parameters such as precipitation, temperature, humidity, and wind speed, was normalized to a uniform distribution range of 0–1. The time series data was divided into two parts: an 80% training set and a 20% validation set. The training set was used for network training, and the validation set was used for model evaluation.

[0103] An optimal prediction network architecture was constructed, consisting of two main modules: an encoder and a decoder. The encoder uses a multi-head self-attention mechanism with eight attention heads, each with a dimension of 64. Attention scores are calculated by calculating the dot product similarity between feature vectors, and attention weights are obtained after softmax normalization. The attention weights are multiplied and concatenated with the feature vectors to obtain a weighted feature representation. The decoder uses a cross-attention mechanism to associate the feature representation output by the encoder with the prediction target to generate the final prediction result.

[0104] During the network training phase, the Adam optimizer was used for parameter optimization, with an initial learning rate of 0.001. If the validation loss did not decrease for five consecutive epochs, the learning rate was reduced to 0.1 times the original value. A gradient clipping threshold of 5.0 was set to prevent gradient explosion. Training was terminated when the mean absolute error on the validation set fell below 0.1. Using precipitation data from January to December 2020 for a specific city as an example, after 100 rounds of training, the validation set error dropped to 0.08, indicating a fully trained prediction network.

[0105] A Monte Carlo search tree is constructed to explore the parameter space. The root node of the tree stores the current optimal network parameter state. Each time a node is expanded, the value range of a parameter is randomly adjusted to generate multiple candidate solutions. The candidate solutions are input into the trained prediction network to calculate the prediction error, and the top five solutions with the lowest error are selected as child nodes. Each child node records the parameter value and corresponding score.

[0106] After the search tree is constructed, a backtracking update is performed. Starting from the leaf nodes and propagating upward, the scores of child nodes are calculated based on the inverse of the prediction error. Nodes with more visits receive a larger exploration bonus coefficient. After 1000 sampling iterations, the visit frequency and score distribution of the nodes corresponding to each parameter value are calculated. The visit frequency is normalized to obtain the exploration degree index, and the mean of the score distribution is used as the importance index. The weighted average of these two indicators is used to quantify the parameter's influence.

[0107] Generate a parameter sensitivity distribution map. Plot a curve with the parameter value on the horizontal axis and the quantitative impact value on the vertical axis. Peaks on the curve indicate a significant impact on the forecast, while troughs indicate a smaller impact. This map allows for intuitive analysis of the sensitivity of different parameters, providing a basis for decision-making in precipitation control.

[0108] In this embodiment, the dependencies between features are adaptively extracted through the self-attention mechanism, which avoids the subjectivity brought by manually setting feature weights and improves the accuracy of feature representation. The parameter space is intelligently sampled in combination with the Monte Carlo tree search method. Compared with traditional grid search and random search methods, it can more efficiently explore important areas in the parameter space and reduce computational overhead. The degree of parameter influence is quantified by combining access frequency and score distribution, which not only takes into account the adequacy of the exploration of parameter values, but also reflects the contribution of parameters to prediction performance, making the parameter sensitivity analysis results more comprehensive and reliable.

[0109] S2. Execute a precipitation forecast simulation based on the first-generation candidate parameter combination to generate precipitation output data, extract precipitation features, and calculate channel dependencies on the extracted feature graph to obtain a precipitation spatial feature graph. Extract precipitation evolution patterns to obtain a temporal feature sequence. Model the causal relationship between the temporal feature sequences using a causal convolutional network, generate a temporal causal graph, and calculate structural similarity using a graph matching network to obtain a temporal association matrix. Add the precipitation spatial feature graph, the temporal association matrix, and the precipitation control feature tensor to a feature fusion module based on a capsule network to generate a precipitation feature vector. Calculate the energy difference between the precipitation feature vector and the measured precipitation data using an energy-based contrastive learning framework to obtain feature matching, and score the forecast effect using a hierarchical fluctuation assessment network. Adaptively fuse the scoring results for different intensity intervals to obtain a comprehensive forecast score.

[0110] The channel dependency refers to the correlation or interaction between channels in the feature channel, which is used to optimize the feature representation or model structure. The graph matching network is a network model that achieves inter-graph matching by comparing the features of the nodes and edges of the graph, and is widely used in graph comparison and graph search tasks. The time series association matrix is a matrix that describes the relationship between different time points, and is used to analyze the dynamic characteristics and patterns of time series data. The feature fusion module based on the capsule network is a module that uses the capsule network to capture the complex relationship between features and fuse them, which is suitable for multimodal data processing. The energy-based contrastive learning framework is a contrastive learning method that measures the differences between samples through energy functions and optimizes the model, which is used to improve feature differentiation capabilities. The hierarchical fluctuation evaluation network is a network that analyzes data fluctuation characteristics and evaluates their impact through a hierarchical structure, and is used for time series data modeling.

[0111] In an optional embodiment,

[0112] Based on the first generation of candidate parameter combinations, precipitation forecast simulation is performed to generate precipitation output data, precipitation features are extracted and channel dependencies are calculated for the extracted feature graphs to obtain precipitation spatial feature graphs and precipitation evolution laws are extracted to obtain time series feature sequences. The causal relationship between the time series feature sequences is modeled through a causal convolutional network, a time series causal graph is generated and structural similarity is calculated through a graph matching network to obtain a time series association matrix. The precipitation spatial feature graph, the time series association matrix and the precipitation control feature tensor are added to a feature fusion module based on a capsule network to generate a precipitation feature vector. The energy difference between the precipitation feature vector and the measured precipitation data is calculated through an energy-based contrastive learning framework to obtain feature matching and score the forecast effect through a hierarchical fluctuation evaluation network. The scoring results of different intensity intervals are adaptively fused to obtain a comprehensive forecast score including:

[0113] constructing a test parameter combination input matrix based on the first generation candidate parameter combinations, performing normalization processing on the test parameter combination input matrix to obtain a normalized parameter matrix, and inputting the normalized parameter matrix into a preset numerical forecast model to perform precipitation forecast simulation and generate precipitation output data;

[0114] Constructing multiple parallel feature extraction branches to extract precipitation features from the precipitation output data, wherein the feature extraction branches extract precipitation features of different scales through a multi-layer convolutional network, perform batch normalization and activation function processing after each convolution layer, and reduce the feature dimension through maximum pooling to obtain a feature map;

[0115] Constructing a channel attention module to calculate the channel dependency of the feature map, compressing the feature map through global average pooling to obtain a channel descriptor, processing the channel descriptor using a multi-layer fully connected network to obtain a channel weight, and reweighting the feature map based on the channel weight to obtain a precipitation spatial feature map;

[0116] The precipitation field features of multiple forecast times are connected in series to construct a time series feature sequence. The precipitation evolution law is extracted through a multi-layer dilated causal convolutional network. The dilated causal convolutional network expands the receptive field by increasing the dilation rate layer by layer and adds residual connections to prevent gradient vanishing. The time series features are constructed into a time series causal graph. The causal relationship strength between nodes in the time series causal graph is calculated based on the attention mechanism. The measured precipitation data is constructed into a reference causal graph. The structural similarity between the time series causal graph and the reference causal graph is calculated using a graph matching network to obtain a time series correlation matrix.

[0117] Constructing a feature fusion module based on a capsule network, the feature fusion module includes multiple primary capsule layers and advanced capsule layers. The primary capsule layers process the precipitation spatial feature map, the temporal correlation matrix, and the precipitation control feature tensor respectively, and integrate multi-source features in the advanced capsule layer through a dynamic routing algorithm to obtain a precipitation feature vector;

[0118] An energy-based contrastive learning framework is constructed. The precipitation feature vector and the measured precipitation data are input into a feature encoder and mapped to a contrastive learning space. Positive and negative sample pairs are constructed in preset batches. The feature matching degree is calculated based on the energy difference between the positive and negative sample pairs. A hierarchical fluctuation assessment network is constructed. The hierarchical fluctuation assessment network includes multiple evaluation branches for different precipitation intensity intervals. Each evaluation branch extracts the forecast features of the corresponding intensity interval through a multi-layer convolutional network and an attention mechanism to obtain a branch score. The feature matching degree is used as a weight adjustment factor for each branch score to fuse the branch scores to obtain a comprehensive forecast score.

[0119] The dilated causal convolutional network is a convolutional network under causal constraints, which captures long-term dependencies through dilated convolution and is suitable for time series prediction and generation tasks.

[0120] Construct a test parameter combination input matrix. This matrix contains meteorological parameters such as temperature, humidity, and air pressure. The matrix dimensions are 24 × 10 × 10, where 24 represents the time step and 10 × 10 represents the spatial grid distribution. Normalize the input matrix to map each parameter value to the range 0–1. This normalized parameter matrix is then input into the WRF numerical forecast model to perform precipitation forecast simulations and generate precipitation output data.

[0121] When extracting features from precipitation output data, three parallel feature extraction branches are constructed. Each branch contains four convolutional layers with kernel sizes of 3×3, 5×5, and 7×7, respectively, to extract precipitation features at different scales. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function, and feature dimensionality is reduced through 2×2 max pooling. For example, in the first branch, the input data dimensions are 24×10×10, and after four layers of convolution-pooling, the resulting feature map dimensions are 24×5×5.

[0122] When constructing the channel attention module, global average pooling is performed on the feature map to obtain channel descriptors. These channel descriptors are passed through a two-layer fully connected network with 128 hidden layer nodes and an output layer with the same number of channels. The output is mapped to the range 0-1 using a sigmoid function to obtain channel weights. The channel weights are then multiplied by the original feature map to achieve feature reweighting, resulting in a precipitation spatial feature map.

[0123] To extract temporal features, precipitation field features from 24 consecutive time periods are concatenated to construct a temporal feature sequence. A three-layer dilated causal convolutional network is used, with dilation rates of 1, 2, and 4, and a convolution kernel size of 3. Residual connections are added after each convolution layer to prevent gradient vanishing due to network deepening. The resulting temporal features are constructed into a temporal causal graph, where nodes represent feature states at different moments and edges represent temporal dependencies. A multi-head attention mechanism is used to calculate the strength of causal relationships between nodes. Measured precipitation data is also constructed as a reference causal graph, and a graph matching network is used to calculate the structural similarity between the two graphs to obtain a temporal correlation matrix.

[0124] The feature fusion module consists of three primary capsule layers and one advanced capsule layer. The primary capsule layers process the precipitation spatial feature map, the temporal correlation matrix, and the precipitation control feature tensor, respectively. The output dimension of each primary capsule is 16. A dynamic routing algorithm is used to integrate multi-source features in the advanced capsule layer with three routing iterations, ultimately resulting in a 32-dimensional precipitation feature vector.

[0125] In an energy-based contrastive learning framework, a feature encoder maps precipitation feature vectors and measured precipitation data into a 128-dimensional contrastive learning space. Within each training batch, 32 pairs of positive samples and 96 pairs of negative samples are randomly sampled, and feature matching is calculated based on the energy difference between the pairs. A hierarchical fluctuation assessment network consists of three evaluation branches, one for light rain, one for moderate rain, and one for heavy rain. Each branch consists of two convolutional layers and a channel attention module to extract forecast features for the corresponding intensity range and generate a branch score. The branch scores are weighted and averaged using the feature matching as a weighting factor to obtain the final comprehensive forecast score.

[0126] In this embodiment, through the multi-scale feature extraction branch and channel attention mechanism, the spatial distribution characteristics of the precipitation field can be effectively captured, and the spatial refinement of the precipitation forecast can be improved. The temporal feature extraction method based on the dilated causal convolutional network and graph matching can accurately characterize the precipitation evolution law and enhance the temporal continuity of the forecast results. The capsule network is used for feature fusion and energy-based contrast learning is introduced to achieve adaptive integration of multi-source features and improve the forecast accuracy of different precipitation intensity ranges.

[0127] In an optional embodiment,

[0128] Construct a feature fusion module based on a capsule network. The feature fusion module includes multiple primary capsule layers and advanced capsule layers. The primary capsule layers process the precipitation spatial feature map, the temporal correlation matrix, and the precipitation control feature tensor respectively. The dynamic routing algorithm is used to integrate multi-source features in the advanced capsule layer to obtain a precipitation feature vector. The method includes:

[0129] Constructing a feature fusion module based on a capsule network, the feature fusion module includes multiple primary capsule layers and multiple high-level capsule layers. The primary capsule layers respectively process a precipitation spatial feature map, a temporal association matrix, and a precipitation control feature tensor. The primary capsule layers include multiple capsule units with fixed dimensions. Feature channels of the precipitation spatial feature map are reorganized into multiple groups and mapped to a capsule space through a feature transformation matrix to obtain a first group of capsule representations. The temporal association matrix is divided into multiple overlapping sub-matrices and mapped to a capsule space through a feature transformation matrix to obtain a second group of capsule representations. The precipitation control feature tensor is grouped along a physical quantity dimension and mapped to a capsule space through a feature transformation matrix to obtain a third group of capsule representations.

[0130] The advanced capsule layer includes a plurality of capsule units of fixed dimensions, and a coupling coefficient matrix is initialized to establish a connection relationship between the primary capsule layer and the advanced capsule layer. The dimension of the coupling coefficient matrix is the number of primary capsules multiplied by the number of advanced capsules;

[0131] Mapping the capsule representation of the primary capsule layer into a prediction vector through a feature transformation matrix, calculating a temporary representation of the high-level capsule layer based on the prediction vector and the coupling coefficient matrix, and performing nonlinear compression on the temporary representation to obtain an activation representation of the high-level capsule layer;

[0132] The similarity of the predicted vector is calculated based on the activation representation of the high-level capsule layer, the coupling coefficient in the coupling coefficient matrix is updated according to the similarity and normalization is performed, the connection relationship between the primary capsule layer and the high-level capsule layer is repeatedly optimized through multiple rounds of iterations using a dynamic routing algorithm, and the temporary representation and activation representation of the high-level capsule layer are sequentially connected based on the connection relationship to generate a precipitation feature vector, wherein the precipitation feature vector includes precipitation spatial distribution information, temporal evolution information, and physical control information.

[0133] The capsule unit is the basic component of the capsule network, which is used to capture the spatial relationship and position information between features and enhance the expressiveness of the model. The overlapping sub-matrix is a sub-block with overlapping areas extracted from the feature matrix, which is used to refine feature analysis or structural optimization. The coupling coefficient matrix is a matrix that describes the degree of coupling between features or nodes and is used to quantify the interaction strength. The activation representation is a feature representation used to express the activation state of neurons in the neural network, reflecting the degree of influence of the input on the network. The dynamic routing algorithm is an algorithm used in the capsule network to determine the dependency relationship between capsules, and the feature transfer of the optimal path is achieved through iterative adjustment.

[0134] Construct the overall architecture of the feature fusion module. This module consists of three primary capsule layers and two advanced capsule layers. The primary capsule layers process three types of input data: the precipitation spatial feature map, the temporal correlation matrix, and the precipitation control feature tensor. Each primary capsule layer contains 32 capsule units with a dimension of 16. The advanced capsule layers contain 16 capsule units with a dimension of 32.

[0135] To process the precipitation spatial feature map, its 128 feature channels are first reorganized into 32 groups of 4 channels each. Each feature group is mapped to a 16-dimensional capsule space using a 4×16 feature transformation matrix, resulting in the first set of capsule representations. In practice, the precipitation feature map is reconstructed into a 32×4 matrix and multiplied by the transformation matrix to obtain a 32×16 capsule representation.

[0136] To process the temporal correlation matrix, we divide it into 32 overlapping submatrices, each 4×4 in size. Similarly, each submatrix is mapped to the capsule space using a 16×16 feature transformation matrix, resulting in a second set of capsule representations. The overlapping region between submatrices is set to two time steps to ensure the continuity of temporal information.

[0137] The precipitation control feature tensor is divided into 32 groups along the physical quantity dimension, each containing four physical quantities. A 4×16 feature transformation matrix is used to map it to the capsule space, yielding a third set of capsule representations. Physical quantities can be grouped based on their physical relevance, for example, related quantities such as temperature and humidity can be grouped together.

[0138] Establish connections between the primary capsule layer and the higher-level capsule layer. Initialize a 96×16 coupling coefficient matrix, where 96 is the total number of capsules in the three primary capsule layers (32 per layer) and 16 is the number of capsules in the higher-level capsule layer. Initialize the coupling coefficients uniformly between 0 and 1, and normalize each row so that they sum to 1.

[0139] The 16-dimensional representation of the primary capsule is mapped to a 32-dimensional prediction vector using a 16×32 feature transformation matrix. A temporary representation of the higher-level capsule layer is calculated based on the prediction vector and the coupling coefficient. The prediction vector is multiplied by the corresponding coupling coefficient and the sum is calculated. The temporary representation is then nonlinearly compressed using the squash function to obtain the activation representation.

[0140] During the dynamic routing iteration, the cosine similarity between the prediction vector and the high-level capsule activation representation is calculated, and the coupling coefficient is updated based on this similarity. After each update, the coupling coefficient is softmax-normalized. After three iterations, the optimized connection relationship is obtained. The temporary representation and activation representation of the high-level capsule layer are concatenated along the feature dimension to obtain a 64-dimensional precipitation feature vector.

[0141] In this embodiment, the capsule network is used to effectively integrate multi-source heterogeneous features related to precipitation prediction, retaining the spatial structure information and entity integrity of the features and improving the robustness of feature representation. A dynamic routing mechanism is used to establish an adaptive connection relationship between primary capsules and advanced capsules, which can automatically adjust the contribution weights of different features according to their importance, thereby improving the flexibility and accuracy of feature fusion. The spatial features, temporal features, and physical control features of precipitation are uniformly mapped to the capsule space for fusion, realizing collaborative modeling of multi-scale features and enhancing the expression ability of the precipitation prediction model for complex weather systems.

[0142] S3. Construct an optimization objective function based on a deep neural network, construct a multi-objective fitness function based on uncertainty according to the comprehensive forecast score and model the uncertainty of the optimization objective function through a Gaussian process, construct a dynamic crossover operator based on the Neural Ordinary Differential Equation and construct a parameter evolution trajectory, determine the crossover position and probability based on the parameter evolution trajectory, construct a search strategy network according to an offline reinforcement learning algorithm, train the strategy network according to the historical parameter optimization trajectory and perform crossover operations on parameter combinations with fitness values higher than a preset fitness threshold to generate a second-generation parameter combination, execute the parameter search guided by the strategy network on the second-generation parameter combination to obtain an optimized parameter combination, calculate the stability of the optimized parameter combination in combination with a variational inference algorithm, and select the optimized parameter combination with the highest stability as the optimal parameter combination output.

[0143] The uncertainty-based multi-objective fitness function is a function that uses uncertainty to quantify the pros and cons of objectives in multi-objective optimization, which helps to balance multiple objectives during the optimization process. The Gaussian process is a non-parametric machine learning method used to model the probability distribution of complex functions and is widely used in prediction and uncertainty quantification. The dynamic crossover operator based on the neural ordinary differential equation is an operator that uses neural ordinary differential equations to model dynamic change rules and is used to optimize the crossover operation in the genetic algorithm. The offline reinforcement learning algorithm is a reinforcement learning method that trains strategies on existing data sets and is suitable for task environments where real-time interaction is not possible.

[0144] In an optional embodiment,

[0145] An optimization objective function is constructed based on a deep neural network, a multi-objective fitness function based on uncertainty is constructed according to the comprehensive forecast score, and the uncertainty of the optimization objective function is modeled through a Gaussian process, a dynamic crossover operator based on a neural ordinary differential equation is constructed and a parameter evolution trajectory is constructed, a crossover position and probability are determined based on the parameter evolution trajectory, a search strategy network is constructed according to an offline reinforcement learning algorithm, the strategy network is trained according to a historical parameter optimization trajectory, and a crossover operation is performed on a parameter combination whose fitness value is higher than a preset fitness threshold to generate a second-generation parameter combination, a parameter search guided by the strategy network is performed on the second-generation parameter combination to obtain an optimized parameter combination, the stability of the optimized parameter combination is calculated in combination with a variational inference algorithm, and the optimized parameter combination with the highest stability is selected as the optimal parameter combination output, which includes:

[0146] Constructing an optimization objective function based on a deep neural network, wherein the deep neural network consists of an input layer, a hidden layer, and an output layer, inputting the comprehensive forecast score into the input layer, mapping the outputs of the input layer neurons to the hidden layer via a weight matrix, performing a rectified linear activation operation on the outputs of the hidden layer to introduce nonlinear characteristics, and mapping the activated features to the output layer via a fully connected layer to generate an objective function value in scalar form;

[0147] Constructing an uncertainty-based multi-objective fitness function based on the comprehensive forecast score, calculating the similarity between sample points using a radial basis kernel function to establish a Gaussian process covariance matrix, iteratively optimizing the signal variance and length scale hyperparameters of the Gaussian process using a maximum likelihood estimation method, calculating the fitness expectation value and uncertainty index of the parameter configuration to be evaluated using the Gaussian process model, and weightedly combining the fitness expectation value and the uncertainty index to generate a multi-objective fitness function value;

[0148] A dynamic crossover operator based on a neural network of ordinary differential equations is constructed, and a multilayer perceptron network is designed as a parameter change rate calculation unit. The multilayer perceptron network outputs the rate of change of the parameter over time by subjecting the current state of the parameter to multi-layer nonlinear transformations. The Runge-Kutta algorithm is used to numerically solve the neural network of ordinary differential equations to obtain a parameter evolution trajectory. The curvature value of the parameter evolution trajectory at each time point is calculated based on the time series data of the parameter evolution trajectory. The time point where the curvature value is greater than a preset curvature threshold is determined as the crossover position. The corresponding crossover probability is set according to the curvature value, wherein the position with a larger curvature value is assigned a higher crossover probability.

[0149] A search strategy network is constructed based on an offline reinforcement learning algorithm. A dual network structure consisting of a strategy network and a value network is designed. The strategy network uses a multi-layer convolutional structure to extract parameter state features and output crossover probability and mutation strength. The value network uses a fully connected structure to evaluate operation value. State transition sequences and reward signals from historical parameter optimization trajectories are collected to construct an experience replay pool. Training data is batch-sampled from the experience replay pool to optimize the network parameters of the dual network structure.

[0150] Inputting a parameter combination whose fitness value is higher than a preset fitness threshold into the policy network, obtaining a crossover probability and a mutation strength, performing a simulated binary crossover operation at a high curvature position of the parameter evolution trajectory to generate a second-generation parameter combination, performing a parameter search guided by the policy network on the second-generation parameter combination, inputting the parameter combination into the policy network to obtain a search direction and a step size, updating the parameters along the search direction and calculating the objective function value, determining whether to accept the parameter update based on a change in the objective function value and an acceptance probability output by the policy network, and performing multiple rounds of iterative search on each set of parameters to obtain an optimized parameter combination;

[0151] A variational distribution in the form of a diagonal Gaussian is constructed in combination with the variational inference algorithm. The optimized parameter combination is input into the variational distribution to calculate the posterior probability. The variational parameters are iteratively optimized using the stochastic gradient descent algorithm to minimize the relative entropy between the variational distribution and the true posterior distribution. The variance of the variational distribution is calculated as a measure of parameter stability, and the optimized parameter combination with the smallest variance is selected as the optimal parameter combination output.

[0152] The maximum likelihood estimation method is a method for estimating model parameters by maximizing the probability of observed data, and is one of the basic methods of statistical modeling. The Runge-Kutta algorithm is a method for numerically solving differential equations, and is often used for high-precision dynamic system simulation and solution. The simulated binary crossover operation is a crossover strategy in a genetic algorithm, which performs reorganization optimization of continuous variables by simulating binary coding operations.

[0153] An optimization objective function based on a deep neural network was constructed. This deep neural network employs a multi-layered structure. The input layer receives comprehensive forecast score data and has eight input neurons. The hidden layer employs a three-layer structure, with 64, 32, and 16 neurons in each layer, respectively. Input features are mapped to the hidden layer space via a weight matrix. A ReLU activation function is applied to the hidden layer output to introduce nonlinearity, with an activation threshold set to 0.01. Finally, a fully connected layer maps the features to the output layer to obtain a scalar objective function value. Taking actual weather forecast score data as an example, eight scoring indicators, including a temperature forecast score of 0.85 and a precipitation forecast score of 0.78, were input. The network calculated the objective function value to be 0.92.

[0154] A multi-objective fitness function based on uncertainty was constructed. The radial basis kernel function was used to calculate the similarity between sample points, with the kernel bandwidth parameter set to 2.0. The signal variance and length scale hyperparameters of the Gaussian process model were iteratively optimized using maximum likelihood estimation, with the initial values of the signal variance set to 1.0, the initial value of the length scale set to 0.5, and the learning rate set to 0.01 for 500 iterations. For each parameter configuration to be evaluated, the expected fitness value and uncertainty index were calculated using the trained Gaussian process model. The expected fitness value and uncertainty index were weighted together with weights of 0.7 and 0.3 to generate the final multi-objective fitness function value.

[0155] A dynamic crossover operator based on the neural ordinary differential equation was constructed. A three-layer perceptron network was designed as the parameter change rate calculation unit, with each hidden layer containing 32 neurons. The network receives the current state of the parameter and outputs the rate of change of the parameter over time through nonlinear transformation. The neural ordinary differential equation was numerically solved using the fourth-order Runge-Kutta algorithm, with a time step of 0.01 and a solution duration of 1.0, to obtain the parameter evolution trajectory. The curvature value of each time point was calculated based on the trajectory time series data, and the curvature threshold was set to 0.5. The time point with a curvature value greater than the threshold was determined as the crossover position. The crossover probability was set according to the curvature value. When the curvature value was 0.6, the crossover probability was set to 0.3, and when the curvature value was 0.8, the crossover probability was set to 0.6.

[0156] Construct a search strategy network. The strategy network uses a three-layer convolutional structure with a kernel size of 3 and channels of 16, 32, and 64, respectively. It extracts parameter state features and outputs crossover probability and mutation strength. The value network uses a three-layer fully connected structure with 128, 64, and 1 neurons, respectively, to evaluate operational value. Historical optimization data containing 1,000 state transition sequences is collected to construct an experience replay pool. The dual network structure is trained using 32 random samples at a time for 10,000 training rounds.

[0157] Parameter combinations with fitness values greater than 0.8 were fed into the policy network to obtain crossover probabilities and mutation strengths. Simulated binary crossover was performed at locations with high curvature in the parameter evolution trajectory, with a crossover distribution index set to 20. A policy network-guided parameter search was then performed on the resulting second-generation parameter combinations. Each iteration, the parameter combination was fed into the policy network to obtain the search direction and step size, with the search step size initially set to 0.1. The parameters were updated along the search direction and the objective function value was calculated. Acceptance of the update was determined based on the improvement in the objective function value and the acceptance probability output by the policy network, with an acceptance probability threshold set to 0.5. 100 rounds of iterative search were performed for each parameter set to obtain the optimal parameter combination.

[0158] Parameter stability was assessed using a variational inference algorithm. A diagonal Gaussian variational distribution was constructed, with the mean initialized to the value of the optimized parameter combination and the variance initialized to 0.1. The variational parameters were optimized using stochastic gradient descent with a learning rate of 0.001. The algorithm was iterated for 2000 rounds to minimize the relative entropy between the variational distribution and the true posterior distribution. The variance of the variational distribution was calculated as a measure of parameter stability, and the optimized parameter combination with the smallest variance was selected as the final output.

[0159] In this embodiment, the optimization objective function constructed by deep neural network can effectively capture the nonlinear relationship between forecast scoring indicators, improve the guiding ability of objective function for parameter optimization, the uncertainty modeling method based on Gaussian process can simultaneously consider the expected value and uncertainty of parameter fitness, and avoid falling into local optimal solution, the dynamic crossover operator based on Neural Ordinary Differential Equation can adaptively determine the crossover position and probability, and improve the efficiency of parameter search, the policy network constructed by offline reinforcement learning can learn search strategies from historical optimization experience, accelerate the parameter optimization process, and the variational inference algorithm can evaluate the stability of the optimization parameter combination, and ensure the robustness of the final output parameters. In summary, this embodiment realizes the intelligence and automation of the parameter optimization process, and significantly improves the optimization effect and efficiency.

[0160] In an optional embodiment,

[0161] A dynamic crossover operator based on the Neural Ordinary Differential Equation is constructed, and a multilayer perceptron network is designed as a parameter change rate calculation unit. The multilayer perceptron network outputs the parameter change rate over time by subjecting the current state of the parameter to multi-layer nonlinear transformations. The Runge-Kutta algorithm is used to numerically solve the Neural Ordinary Differential Equation to obtain the parameter evolution trajectory, including:

[0162] A multilayer perceptron network is designed as a parameter change rate calculation unit. The multilayer perceptron network consists of an input layer, multiple hidden layers, and an output layer. The input layer receives the current state of the parameter. The number of neurons in the multiple hidden layers decreases layer by layer, and a normalization layer is set at the output end of each layer. The normalized features are nonlinearly transformed using a hyperbolic tangent function. The output layer maps the transformed features through a fully connected layer to obtain the rate of change of the parameter over time.

[0163] Extracting a parameter evolution sequence from a historical optimization database, normalizing parameter values in the parameter evolution sequence and calculating parameter changes between adjacent sampling points, using the normalized parameter values as inputs to the multilayer perceptron network, and training the multilayer perceptron network using the parameter changes as supervisory signals, and stopping training when the validation set loss does not decrease for multiple consecutive rounds;

[0164] The trained multilayer perceptron network is embedded in an ordinary differential equation framework to construct a core component of a dynamic crossover operator, wherein the dynamic crossover operator includes a parameter evolution part and a crossover operation part. In the parameter evolution part, the parameter configuration to be crossed is used as the initial state, and the right-hand side of the neural ordinary differential equation is constructed by the rate of change output by the multilayer perceptron network. The neural ordinary differential equation is numerically solved, and the integration interval is divided into an initial step sequence. The parameter state is input into the multilayer perceptron network at different time points to obtain a change rate estimate. The intermediate point prediction and weighted combination of the Runge-Kutta algorithm are used to calculate the parameter state at the next time point. The integration step is adaptively adjusted according to the truncation error between adjacent time points. The neural ordinary differential equation is numerically solved to obtain a candidate parameter evolution trajectory.

[0165] The candidate parameter evolution trajectories are organized in time series, the average rate of change, rate of change variance, and state difference of the candidate parameter evolution trajectories are calculated, and key turning points on the candidate parameter evolution trajectories are identified. Based on the key turning points, the convergence characteristics of the candidate parameter evolution trajectories are evaluated according to the attenuation trend of the parameter change rate, the wandering characteristics in the state space, and the stability of the trajectory endpoint. The unconverged trajectories are re-solved by adjusting the integration interval, and the candidate parameter evolution trajectories are projected into the physical constraint space. The physical rationality of the candidate parameter evolution trajectories is verified through parameter value range inspection, physical correlation evaluation, and simplified physical model.

[0166] The candidate parameter evolution trajectories that pass the convergence test and the physical rationality test are scored for quality, and the parameter evolution trajectories are screened according to the trajectory smoothness, stability and physical interpretability and stored in the optimization experience database.

[0167] The right-hand term refers to the function or term on the right side of the differential equation, which usually represents the external input or influencing factor of the system dynamics. The intermediate point prediction is a technology that predicts the intermediate state of the dynamic system through interpolation or approximate calculation, which is used to improve the resolution or accuracy of the simulation.

[0168] Design a multilayer perceptron network architecture. The network consists of one input layer, three hidden layers, and one output layer. The input layer has the same number of neurons as the parameter dimensions and receives the normalized parameter states. The three hidden layers have decreasing numbers of neurons, 128, 64, and 32, respectively. Each hidden layer is followed by a BatchNormalization layer and a Tanh activation function for feature normalization and nonlinear transformation. The output layer uses a fully connected projection to achieve a parameter change rate that matches the input dimensions.

[0169] Training data was extracted from the historical optimization database. The parameter evolution sequence was normalized to the minimum and maximum values, scaling the range to the interval [-1, 1]. The parameter difference between adjacent sampling points was calculated as the rate of change label. The training and validation sets were split 80% of the time. The network was trained using the Adam optimizer with a learning rate of 0.001 and a batch size of 32. Early stopping was used when the validation set loss did not decrease for five consecutive rounds.

[0170] The trained network is embedded in an ordinary differential equation framework. The parameter configuration to be crossed is used as the initial value, the integration interval is set to [0, 1], and the initial step size is 0.1. At each time point, the parameter state is input into the network to obtain the rate of change. A fourth-order Runge-Kutta algorithm is used for the solution, which consists of four stages: first, the rate of change k1 of the current state is calculated; then, a half-step size and k1 are used to predict an intermediate state and calculate the rate of change k2; then, a half-step size and k2 are used to predict another intermediate state and calculate k3; finally, a full step size and k3 are used to predict the final state and calculate k4. A weighted combination of k1 to k4 is performed to obtain the next state.

[0171] Adaptively adjust the step size based on the truncation error. Calculate the difference between the fourth-order and fifth-order Runge-Kutta solutions to estimate the truncation error. If the error is greater than the upper tolerance limit of 0.001, halve the step size and recalculate. If the error is less than the lower tolerance limit of 0.0001, double the step size. Repeat this process until the integration is complete, obtaining the candidate parameter evolution trajectory.

[0172] Perform convergence analysis on candidate trajectories. Calculate the average rate of change, variance, and state difference between adjacent points on the trajectory. Identify key turning points by observing sudden changes in the rate of change. Evaluate whether the rate of change exhibits an exponential decay trend, whether the trajectory exhibits stable movement in state space, and whether the endpoint parameters converge. For non-convergent trajectories, re-solve by extending the integration interval.

[0173] Perform physical plausibility checks. Project the trajectory back to the original parameter space to check whether it meets the parameter range constraints. Evaluate the rationality of the correlation between parameters based on a simplified physical model, such as whether flow and pressure are positively correlated. Rapidly simulate the trajectory using a reduced-order physical model to verify its physical feasibility.

[0174] The trajectory quality is scored based on three dimensions: smoothness (degree of state transition between adjacent points), stability (endpoint convergence characteristics), and physical interpretability (conformity with physical laws). The trajectory with the highest overall score is selected as the final evolution trajectory and stored in the optimization experience library.

[0175] In this embodiment, the parameter evolution law is learned through a multi-layer perceptron network, and the discrete optimization experience is converted into a continuous dynamic system, which realizes the intelligence and automation of the parameter optimization process and improves the optimization efficiency. The numerical solution method with adaptive step size and multiple verification mechanism are adopted to ensure the convergence and physical rationality of the parameter evolution trajectory, enhance the reliability and robustness of the optimization results, construct a parameter evolution model based on historical data, make full use of existing optimization experience, and realize the inheritance of optimization knowledge and continuous improvement of optimization capabilities through continuous accumulation and updating of the optimization database.

[0176] Figure 2 FIG. 1 is a schematic diagram of the structure of a mesoscale convection parameter optimization system based on a genetic algorithm according to an embodiment of the present invention. Figure 2 As shown, the system includes:

[0177] The first unit is used to perform dynamic system modeling of the historical precipitation process through a preset neural ordinary differential equation to obtain a precipitation evolution state equation, identify dynamic system parameters in combination with a variational inference method and map them to the convection parameter space to obtain parameter initial thresholds, extract multidimensional features of precipitation data and atmospheric environment data through a spatiotemporal heterogeneous graph neural network, perform denoising on the extracted features to generate parameter influence features through a diffusion probability model, extract physical constraints of the convection process from the parameter influence features through a physics-guided neural network and correct the parameter initial thresholds, combine the cross-modal visual attention network to fuse the corrected parameter initial thresholds and the precipitation evolution sequence corresponding to the precipitation evolution state equation to obtain a precipitation control feature tensor and train an optimal prediction network, generate a parameter sensitivity distribution map based on the optimal prediction network and Monte Carlo tree search sampling, perform spectral clustering to construct a parameter initialization manifold, and generate a first-generation candidate parameter combination in combination with a meta-learning-driven genetic algorithm;

[0178] The second unit is used to perform precipitation forecast simulation based on the first-generation candidate parameter combination to generate precipitation output data, extract precipitation features and calculate channel dependencies on the extracted feature graph to obtain a precipitation spatial feature graph and extract precipitation evolution laws to obtain a time series feature sequence, model the causal relationship between the time series feature sequences through a causal convolutional network, generate a time series causal graph and calculate structural similarity through a graph matching network to obtain a time series association matrix, add the precipitation spatial feature graph, the time series association matrix and the precipitation control feature tensor to a feature fusion module based on a capsule network to generate a precipitation feature vector, calculate the energy difference between the precipitation feature vector and the measured precipitation data through an energy-based contrastive learning framework, obtain feature matching, and score the forecast effect through a hierarchical fluctuation evaluation network, and adaptively fuse the scoring results of different intensity intervals to obtain a comprehensive forecast score;

[0179] The third unit is used to construct an optimization objective function based on a deep neural network, construct a multi-objective fitness function based on uncertainty according to the comprehensive forecast score and model the uncertainty of the optimization objective function through a Gaussian process, construct a dynamic crossover operator based on a neural ordinary differential equation and construct a parameter evolution trajectory, determine the crossover position and probability based on the parameter evolution trajectory, construct a search strategy network according to an offline reinforcement learning algorithm, train the strategy network according to the historical parameter optimization trajectory and perform a crossover operation on the parameter combination whose fitness value is higher than a preset fitness threshold to generate a second-generation parameter combination, perform a parameter search guided by the strategy network on the second-generation parameter combination to obtain an optimized parameter combination, calculate the stability of the optimized parameter combination in combination with a variational inference algorithm, and select the optimized parameter combination with the highest stability as the optimal parameter combination output.

[0180] According to a third aspect of the embodiments of the present invention,

[0181] An electronic device is provided, comprising:

[0182] processor;

[0183] a memory for storing processor-executable instructions;

[0184] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0185] According to a fourth aspect of the embodiments of the present invention,

[0186] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0187] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0188] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A mesoscale convection parameter optimization method based on genetic algorithm, characterized in that: include: The historical precipitation process is modeled by the dynamic system using a pre-set neural ordinary differential equation to obtain the precipitation evolution state equation. The variational inference method is used to identify the dynamic system parameters and map them to the convection parameter space to obtain the parameter initial threshold. The multidimensional features of precipitation data and atmospheric environment data are extracted through a spatiotemporal heterogeneous graph neural network. The extracted features are denoised using a diffusion probability model to generate parameter influence features. The physical constraints of the convection process are extracted from the parameter influence features through a physics-guided neural network and the parameter initial threshold is corrected. The corrected parameter initial threshold is fused with the precipitation evolution sequence corresponding to the precipitation evolution state equation using a cross-modal visual attention network to obtain the precipitation control feature tensor and train the optimal prediction network. Based on the optimal prediction network and Monte Carlo tree search sampling, a parameter sensitivity distribution map is generated. Spectral clustering is performed to construct a parameter initialization manifold and a meta-learning-driven genetic algorithm is used to generate the first generation of candidate parameter combinations. Based on the first generation of candidate parameter combinations, a precipitation forecast simulation is performed to generate precipitation output data, precipitation features are extracted and channel dependencies are calculated for the extracted feature graph to obtain a precipitation spatial feature graph and precipitation evolution law is extracted to obtain a time series feature sequence, the causal relationship between the time series feature sequences is modeled through a causal convolutional network, a time series causal graph is generated and structural similarity is calculated through a graph matching network to obtain a time series association matrix, the precipitation spatial feature graph, the time series association matrix and the precipitation control feature tensor are added to a feature fusion module based on a capsule network to generate a precipitation feature vector, the energy difference between the precipitation feature vector and the measured precipitation data is calculated through an energy-based contrastive learning framework to obtain feature matching, and the forecast effect is scored through a hierarchical fluctuation evaluation network, and the scoring results of different intensity intervals are adaptively fused to obtain a comprehensive forecast score; An optimization objective function is constructed based on a deep neural network, a multi-objective fitness function based on uncertainty is constructed according to the comprehensive forecast score, and the uncertainty of the optimization objective function is modeled through a Gaussian process, a dynamic crossover operator based on a neural ordinary differential equation is constructed and a parameter evolution trajectory is constructed, the crossover position and probability are determined based on the parameter evolution trajectory, a search strategy network is constructed according to an offline reinforcement learning algorithm, the strategy network is trained according to the historical parameter optimization trajectory, and a crossover operation is performed on the parameter combination whose fitness value is higher than a preset fitness threshold to generate a second-generation parameter combination, a parameter search guided by the strategy network is performed on the second-generation parameter combination to obtain an optimized parameter combination, the stability of the optimized parameter combination is calculated in combination with a variational inference algorithm, and the optimized parameter combination with the highest stability is selected as the optimal parameter combination output.

2. The method according to claim 1, characterized in that The historical precipitation process is modeled by the dynamic system using a pre-set neural ordinary differential equation to obtain the precipitation evolution state equation. The variational inference method is used to identify the dynamic system parameters and map them to the convection parameter space to obtain the parameter initial threshold. The multidimensional features of precipitation data and atmospheric environment data are extracted through a spatiotemporal heterogeneous graph neural network. The extracted features are denoised using a diffusion probability model to generate parameter influence features. The physical constraints of the convection process are extracted from the parameter influence features through a physics-guided neural network and the parameter initial threshold is corrected. The modified parameter initial threshold is fused with the precipitation evolution sequence corresponding to the precipitation evolution state equation using a cross-modal visual attention network to obtain the precipitation control feature tensor and train the optimal prediction network. Based on the optimal prediction network and Monte Carlo tree search sampling, a parameter sensitivity distribution map is generated. Spectral clustering is performed to construct a parameter initialization manifold and a meta-learning-driven genetic algorithm is used to generate the first generation of candidate parameter combinations, including: A dynamic system model is performed on the historical precipitation process using a preset Neural Ordinary Differential Equation. The historical precipitation process includes precipitation intensity, precipitation duration, and precipitation spatial distribution data. The historical precipitation process data is input into the Neural Ordinary Differential Equation in time series. An adaptive step-size solver is used to perform time series integration on the historical precipitation process data and fit the dynamic evolution trend. The precipitation evolution state equation that describes the temporal evolution characteristics of precipitation intensity is output; Inputting the precipitation evolution state equation into a variational inference method for parameter extraction, encoding and mapping the precipitation evolution state equation into a parameter distribution space to obtain dynamic system parameters, decoding and reconstructing the dynamic system parameters into a state equation, constructing an optimization target by calculating the error between the reconstructed state equation and the original state equation, and the difference between the parameter distribution and the prior distribution, iterating the optimization until the error converges to extract the dynamic system parameters, and converting the dynamic system parameters into a convection parameter space through physical mapping to obtain an initial parameter threshold; Collect precipitation data and atmospheric environment data, including temperature, relative humidity, and wind field elements. Divide the forecast area into grid points according to a preset spatial resolution. Construct a graph structure based on the grid points. The nodes of the graph structure store the precipitation data and atmospheric environment data corresponding to the grid point positions. The edges of the graph structure calculate the spatial correlation strength based on the spatial distance between the grid points and the correlation with meteorological elements. Input the graph structure into a preset spatiotemporal heterogeneous graph neural network. Node and edge features are extracted through graph convolution operations and a message passing mechanism. The time dimension information is then integrated to obtain multidimensional features. Denoising the multidimensional features using a diffusion probability model, gradually adding Gaussian noise to the features according to a preset step size during a forward diffusion process, wherein the noise variance increases exponentially with the number of diffusion steps, inputting the noisy feature sequence into a pre-set denoising network, and gradually recovering the feature information through a reverse diffusion process, wherein the denoising network retains multi-scale feature details during the reverse diffusion process, and ultimately generating parameter influence features reflecting the mechanism of the parameter's influence on precipitation; Inputting the parameter influence characteristics into a pre-set physics-guided neural network to extract physical constraints of the convection process, wherein the physical constraints include convection triggering conditions, convection development conditions, and parameter validity conditions; and evaluating the physical rationality of the initial parameter thresholds based on the physical constraints of the convection process and performing corrections; The modified parameter initial threshold and the precipitation evolution sequence corresponding to the precipitation evolution state equation are input into a pre-set cross-modal visual attention network, the internal correlation of the parameters is calculated through the self-attention mechanism, and the correlation between the parameters and the precipitation evolution sequence is calculated through the cross-attention mechanism. The parameter information and the precipitation evolution information are projected into a unified feature space and fused with the attention weight to generate a precipitation control feature tensor containing the parameter control effect and the precipitation evolution characteristics; An optimal prediction network is obtained based on the precipitation control feature tensor training, the precipitation control feature tensor is input into the optimal prediction network to extract feature dependencies, and feature sampling is performed in combination with the Monte Carlo tree search method. The feature sampling explores the parameter space through tree node expansion, calculates the node value through backtracking update, and generates a parameter sensitivity distribution map describing the degree of parameter influence based on the sampling results; Spectral clustering is performed on the parameter sensitivity distribution map, an affinity matrix is constructed based on the map data, the eigenvalues and eigenvectors of the affinity matrix are calculated, the parameter sensitivity distribution is projected into the characteristic subspace for clustering analysis, and a parameter initialization manifold describing the geometric structure of the parameter distribution is constructed based on the clustering results; within the parameter initialization manifold, a meta-learning-driven genetic algorithm is used to learn historical optimization experience and quickly adapt to the current parameter search task, and the first generation of candidate parameter combinations that meet the physical meaning are generated in combination with the parameter constraints.

3. The method according to claim 2, characterized in that An optimal prediction network is obtained based on the precipitation control feature tensor training, the precipitation control feature tensor is input into the optimal prediction network to extract feature dependencies, and feature sampling is performed in combination with the Monte Carlo tree search method. The feature sampling explores the parameter space through tree node expansion, calculates the node value through backtracking update, and generates a parameter sensitivity distribution map describing the degree of parameter influence based on the sampling results, including: The precipitation control feature tensor is divided into a training set and a validation set according to a time window, the training set is input into a preset optimal prediction network for training, an encoder of the optimal prediction network is controlled to calculate the similarity between feature vectors through a self-attention mechanism to obtain attention weights, features are weightedly aggregated based on the attention weights to obtain feature dependencies, and a decoder of the optimal prediction network is controlled to construct a mapping relationship between features and prediction targets through a cross-attention mechanism; An initial learning rate is set to 0.001 and the optimal prediction network is optimized. When the verification loss on the verification set does not decrease for five consecutive iterations, the learning rate is adjusted to 0.1 times the original value, the gradient norm is limited to a preset gradient norm threshold, and gradient clipping is performed. When the prediction error on the verification set is monitored to be less than a preset error threshold, the network training is stopped to obtain the current parameter state. Constructing a Monte Carlo search tree and initializing a root node, writing the current parameter state into the root node, inputting the parameter state into the trained optimal prediction network to perform evaluation, expanding and exploring the parameter space through tree nodes, screening multiple parameter adjustment solutions with the highest scores based on the evaluation results to construct child nodes, and writing new parameter values and state scores into the child nodes; Starting from the leaf node of the search tree, perform a backtracking update of the node score along the search path upward, update the score of the current child node according to the score and the number of visits of the child node, and assign exploration rewards to the child node according to the number of visits; Based on the sampling results, a statistical analysis is performed on the number of visits and score distributions of sub-nodes corresponding to different parameter values in the search tree, the degree of exploration of the parameter values is determined according to the number of visits, and the importance of the parameter values is determined according to the score distribution. The number of visits and the score distribution are fused to calculate a quantitative index of the parameter influence degree, and the quantitative index is used to generate a parameter sensitivity distribution map describing the parameter influence degree.

4. The method according to claim 1, wherein Based on the first generation of candidate parameter combinations, precipitation forecast simulation is performed to generate precipitation output data, precipitation features are extracted and channel dependencies are calculated for the extracted feature graphs to obtain precipitation spatial feature graphs and precipitation evolution laws are extracted to obtain time series feature sequences. The causal relationship between the time series feature sequences is modeled through a causal convolutional network, a time series causal graph is generated and structural similarity is calculated through a graph matching network to obtain a time series association matrix. The precipitation spatial feature graph, the time series association matrix and the precipitation control feature tensor are added to a feature fusion module based on a capsule network to generate a precipitation feature vector. The energy difference between the precipitation feature vector and the measured precipitation data is calculated through an energy-based contrastive learning framework to obtain feature matching and score the forecast effect through a hierarchical fluctuation evaluation network. The scoring results of different intensity intervals are adaptively fused to obtain a comprehensive forecast score including: constructing a test parameter combination input matrix based on the first generation candidate parameter combinations, performing normalization processing on the test parameter combination input matrix to obtain a normalized parameter matrix, and inputting the normalized parameter matrix into a preset numerical forecast model to perform precipitation forecast simulation and generate precipitation output data; Constructing multiple parallel feature extraction branches to extract precipitation features from the precipitation output data, wherein the feature extraction branches extract precipitation features of different scales through a multi-layer convolutional network, perform batch normalization and activation function processing after each convolution layer, and reduce the feature dimension through maximum pooling to obtain a feature map; Constructing a channel attention module to calculate the channel dependency of the feature map, compressing the feature map through global average pooling to obtain a channel descriptor, processing the channel descriptor using a multi-layer fully connected network to obtain a channel weight, and reweighting the feature map based on the channel weight to obtain a precipitation spatial feature map; The precipitation field features of multiple forecast times are connected in series to construct a time series feature sequence. The precipitation evolution law is extracted through a multi-layer dilated causal convolutional network. The dilated causal convolutional network expands the receptive field by increasing the dilation rate layer by layer and adds residual connections to prevent gradient vanishing. The time series features are constructed into a time series causal graph. The causal relationship strength between nodes in the time series causal graph is calculated based on the attention mechanism. The measured precipitation data is constructed into a reference causal graph. The structural similarity between the time series causal graph and the reference causal graph is calculated using a graph matching network to obtain a time series correlation matrix. Constructing a feature fusion module based on a capsule network, the feature fusion module includes multiple primary capsule layers and advanced capsule layers. The primary capsule layers process the precipitation spatial feature map, the temporal correlation matrix, and the precipitation control feature tensor respectively, and integrate multi-source features in the advanced capsule layer through a dynamic routing algorithm to obtain a precipitation feature vector; An energy-based contrastive learning framework is constructed. The precipitation feature vector and the measured precipitation data are input into a feature encoder and mapped to a contrastive learning space. Positive and negative sample pairs are constructed in preset batches. The feature matching degree is calculated based on the energy difference between the positive and negative sample pairs. A hierarchical fluctuation assessment network is constructed. The hierarchical fluctuation assessment network includes multiple evaluation branches for different precipitation intensity intervals. Each evaluation branch extracts the forecast features of the corresponding intensity interval through a multi-layer convolutional network and an attention mechanism to obtain a branch score. The feature matching degree is used as a weight adjustment factor for each branch score to fuse the branch scores to obtain a comprehensive forecast score.

5. The method according to claim 4, characterized in that Construct a feature fusion module based on a capsule network. The feature fusion module includes multiple primary capsule layers and advanced capsule layers. The primary capsule layers process the precipitation spatial feature map, the temporal correlation matrix, and the precipitation control feature tensor respectively. The dynamic routing algorithm is used to integrate multi-source features in the advanced capsule layer to obtain a precipitation feature vector. The method includes: Constructing a feature fusion module based on a capsule network, the feature fusion module includes multiple primary capsule layers and multiple high-level capsule layers. The primary capsule layers respectively process a precipitation spatial feature map, a temporal association matrix, and a precipitation control feature tensor. The primary capsule layers include multiple capsule units with fixed dimensions. Feature channels of the precipitation spatial feature map are reorganized into multiple groups and mapped to a capsule space through a feature transformation matrix to obtain a first group of capsule representations. The temporal association matrix is divided into multiple overlapping sub-matrices and mapped to a capsule space through a feature transformation matrix to obtain a second group of capsule representations. The precipitation control feature tensor is grouped along a physical quantity dimension and mapped to a capsule space through a feature transformation matrix to obtain a third group of capsule representations. The advanced capsule layer includes a plurality of capsule units of fixed dimensions, and a coupling coefficient matrix is initialized to establish a connection relationship between the primary capsule layer and the advanced capsule layer. The dimension of the coupling coefficient matrix is the number of primary capsules multiplied by the number of advanced capsules; Mapping the capsule representation of the primary capsule layer into a prediction vector through a feature transformation matrix, calculating a temporary representation of the high-level capsule layer based on the prediction vector and the coupling coefficient matrix, and performing nonlinear compression on the temporary representation to obtain an activation representation of the high-level capsule layer; The similarity of the predicted vector is calculated based on the activation representation of the high-level capsule layer, the coupling coefficient in the coupling coefficient matrix is updated according to the similarity and normalization is performed, the connection relationship between the primary capsule layer and the high-level capsule layer is repeatedly optimized through multiple rounds of iterations using a dynamic routing algorithm, and the temporary representation and activation representation of the high-level capsule layer are sequentially connected based on the connection relationship to generate a precipitation feature vector, wherein the precipitation feature vector includes precipitation spatial distribution information, temporal evolution information, and physical control information.

6. The method according to claim 1, characterized in that An optimization objective function is constructed based on a deep neural network, a multi-objective fitness function based on uncertainty is constructed according to the comprehensive forecast score, and the uncertainty of the optimization objective function is modeled through a Gaussian process, a dynamic crossover operator based on a neural ordinary differential equation is constructed and a parameter evolution trajectory is constructed, a crossover position and probability are determined based on the parameter evolution trajectory, a search strategy network is constructed according to an offline reinforcement learning algorithm, the strategy network is trained according to a historical parameter optimization trajectory, and a crossover operation is performed on a parameter combination whose fitness value is higher than a preset fitness threshold to generate a second-generation parameter combination, a parameter search guided by the strategy network is performed on the second-generation parameter combination to obtain an optimized parameter combination, the stability of the optimized parameter combination is calculated in combination with a variational inference algorithm, and the optimized parameter combination with the highest stability is selected as the optimal parameter combination output, which includes: Constructing an optimization objective function based on a deep neural network, wherein the deep neural network consists of an input layer, a hidden layer, and an output layer, inputting the comprehensive forecast score into the input layer, mapping the outputs of the input layer neurons to the hidden layer via a weight matrix, performing a rectified linear activation operation on the outputs of the hidden layer to introduce nonlinear characteristics, and mapping the activated features to the output layer via a fully connected layer to generate an objective function value in scalar form; Constructing an uncertainty-based multi-objective fitness function based on the comprehensive forecast score, calculating the similarity between sample points using a radial basis kernel function to establish a Gaussian process covariance matrix, iteratively optimizing the signal variance and length scale hyperparameters of the Gaussian process using a maximum likelihood estimation method, calculating the fitness expectation value and uncertainty index of the parameter configuration to be evaluated using the Gaussian process model, and weightedly combining the fitness expectation value and the uncertainty index to generate a multi-objective fitness function value; A dynamic crossover operator based on a neural network of ordinary differential equations is constructed, and a multilayer perceptron network is designed as a parameter change rate calculation unit. The multilayer perceptron network outputs the rate of change of the parameter over time by subjecting the current state of the parameter to multi-layer nonlinear transformations. The Runge-Kutta algorithm is used to numerically solve the neural network of ordinary differential equations to obtain a parameter evolution trajectory. The curvature value of the parameter evolution trajectory at each time point is calculated based on the time series data of the parameter evolution trajectory. The time point where the curvature value is greater than a preset curvature threshold is determined as the crossover position. The corresponding crossover probability is set according to the curvature value, wherein the position with a larger curvature value is assigned a higher crossover probability. A search strategy network is constructed based on an offline reinforcement learning algorithm. A dual network structure consisting of a strategy network and a value network is designed. The strategy network uses a multi-layer convolutional structure to extract parameter state features and output crossover probability and mutation strength. The value network uses a fully connected structure to evaluate operation value. State transition sequences and reward signals from historical parameter optimization trajectories are collected to construct an experience replay pool. Training data is batch-sampled from the experience replay pool to optimize the network parameters of the dual network structure. Inputting a parameter combination whose fitness value is higher than a preset fitness threshold into the policy network, obtaining a crossover probability and a mutation strength, performing a simulated binary crossover operation at a high curvature position of the parameter evolution trajectory to generate a second-generation parameter combination, performing a parameter search guided by the policy network on the second-generation parameter combination, inputting the parameter combination into the policy network to obtain a search direction and a step size, updating the parameters along the search direction and calculating the objective function value, determining whether to accept the parameter update based on a change in the objective function value and an acceptance probability output by the policy network, and performing multiple rounds of iterative search on each set of parameters to obtain an optimized parameter combination; A variational distribution in the form of a diagonal Gaussian is constructed in combination with the variational inference algorithm. The optimized parameter combination is input into the variational distribution to calculate the posterior probability. The variational parameters are iteratively optimized using the stochastic gradient descent algorithm to minimize the relative entropy between the variational distribution and the true posterior distribution. The variance of the variational distribution is calculated as a measure of parameter stability, and the optimized parameter combination with the smallest variance is selected as the optimal parameter combination output.

7. The method according to claim 6, characterized in that A dynamic crossover operator based on the Neural Ordinary Differential Equation is constructed, and a multilayer perceptron network is designed as a parameter change rate calculation unit. The multilayer perceptron network outputs the parameter change rate over time by subjecting the current state of the parameter to multi-layer nonlinear transformations. The Runge-Kutta algorithm is used to numerically solve the Neural Ordinary Differential Equation to obtain the parameter evolution trajectory, including: A multilayer perceptron network is designed as a parameter change rate calculation unit. The multilayer perceptron network consists of an input layer, multiple hidden layers, and an output layer. The input layer receives the current state of the parameter. The number of neurons in the multiple hidden layers decreases layer by layer, and a normalization layer is set at the output end of each layer. The normalized features are nonlinearly transformed using a hyperbolic tangent function. The output layer maps the transformed features through a fully connected layer to obtain the rate of change of the parameter over time. Extracting a parameter evolution sequence from a historical optimization database, normalizing parameter values in the parameter evolution sequence and calculating parameter changes between adjacent sampling points, using the normalized parameter values as inputs to the multilayer perceptron network, and training the multilayer perceptron network using the parameter changes as supervisory signals, and stopping training when the validation set loss does not decrease for multiple consecutive rounds; The trained multilayer perceptron network is embedded in an ordinary differential equation framework to construct a core component of a dynamic crossover operator, wherein the dynamic crossover operator includes a parameter evolution part and a crossover operation part. In the parameter evolution part, the parameter configuration to be crossed is used as the initial state, and the right-hand side of the neural ordinary differential equation is constructed by the rate of change output by the multilayer perceptron network. The neural ordinary differential equation is numerically solved, and the integration interval is divided into an initial step sequence. The parameter state is input into the multilayer perceptron network at different time points to obtain a change rate estimate. The intermediate point prediction and weighted combination of the Runge-Kutta algorithm are used to calculate the parameter state at the next time point. The integration step is adaptively adjusted according to the truncation error between adjacent time points. The neural ordinary differential equation is numerically solved to obtain a candidate parameter evolution trajectory. The candidate parameter evolution trajectories are organized in time series, the average rate of change, rate of change variance, and state difference of the candidate parameter evolution trajectories are calculated, and key turning points on the candidate parameter evolution trajectories are identified. Based on the key turning points, the convergence characteristics of the candidate parameter evolution trajectories are evaluated according to the attenuation trend of the parameter change rate, the wandering characteristics in the state space, and the stability of the trajectory endpoint. The unconverged trajectories are re-solved by adjusting the integration interval, and the candidate parameter evolution trajectories are projected into the physical constraint space. The physical rationality of the candidate parameter evolution trajectories is verified through parameter value range inspection, physical correlation evaluation, and simplified physical model. The candidate parameter evolution trajectories that pass the convergence test and the physical rationality test are scored for quality, and the parameter evolution trajectories are screened according to the trajectory smoothness, stability and physical interpretability and stored in the optimization experience database.

8. A mesoscale convection parameter optimization system based on a genetic algorithm, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to perform dynamic system modeling of the historical precipitation process through a preset neural ordinary differential equation to obtain a precipitation evolution state equation, identify dynamic system parameters in combination with a variational inference method and map them to the convection parameter space to obtain parameter initial thresholds, extract multidimensional features of precipitation data and atmospheric environment data through a spatiotemporal heterogeneous graph neural network, perform denoising on the extracted features to generate parameter influence features through a diffusion probability model, extract physical constraints of the convection process from the parameter influence features through a physics-guided neural network and correct the parameter initial thresholds, combine the cross-modal visual attention network to fuse the corrected parameter initial thresholds and the precipitation evolution sequence corresponding to the precipitation evolution state equation to obtain a precipitation control feature tensor and train an optimal prediction network, generate a parameter sensitivity distribution map based on the optimal prediction network and Monte Carlo tree search sampling, perform spectral clustering to construct a parameter initialization manifold, and generate a first-generation candidate parameter combination in combination with a meta-learning-driven genetic algorithm; The second unit is used to perform precipitation forecast simulation based on the first-generation candidate parameter combination to generate precipitation output data, extract precipitation features and calculate channel dependencies on the extracted feature graph to obtain a precipitation spatial feature graph and extract precipitation evolution laws to obtain a time series feature sequence, model the causal relationship between the time series feature sequences through a causal convolutional network, generate a time series causal graph and calculate structural similarity through a graph matching network to obtain a time series association matrix, add the precipitation spatial feature graph, the time series association matrix and the precipitation control feature tensor to a feature fusion module based on a capsule network to generate a precipitation feature vector, calculate the energy difference between the precipitation feature vector and the measured precipitation data through an energy-based contrastive learning framework, obtain feature matching, and score the forecast effect through a hierarchical fluctuation evaluation network, and adaptively fuse the scoring results of different intensity intervals to obtain a comprehensive forecast score; The third unit is used to construct an optimization objective function based on a deep neural network, construct a multi-objective fitness function based on uncertainty according to the comprehensive forecast score and model the uncertainty of the optimization objective function through a Gaussian process, construct a dynamic crossover operator based on a neural ordinary differential equation and construct a parameter evolution trajectory, determine the crossover position and probability based on the parameter evolution trajectory, construct a search strategy network according to an offline reinforcement learning algorithm, train the strategy network according to the historical parameter optimization trajectory and perform a crossover operation on the parameter combination whose fitness value is higher than a preset fitness threshold to generate a second-generation parameter combination, perform a parameter search guided by the strategy network on the second-generation parameter combination to obtain an optimized parameter combination, calculate the stability of the optimized parameter combination in combination with a variational inference algorithm, and select the optimized parameter combination with the highest stability as the optimal parameter combination output.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images