A Fast Forecasting Method, System, Device, Medium and Terminal for a Large Meteorological Forecasting Model with Low Computational Resources

Abstract

The present invention belongs to the technical field of weather forecasting, and discloses a rapid weather forecasting method, system, device, medium and terminal for large models under low computing resources. Based on the fine-tuning technology of neural networks and using the integration idea, aiming at the problems that the training of large models with a large number of parameters in the existing large models requires high data and computing resources and is difficult to train, etc., for the meteorological field, a rapid forecasting method for large models under low computing resources is designed to break through the technical barriers of existing large model training and realize the rapid forecasting of an effective large model from scratch under low computing resources. The present invention can achieve the rapid forecasting of large models under low computing resources and avoid the high computing resource requirements of existing large model training. The present invention constructs a large model by integrating multiple small models, and solves the problem of difficult training of large models with a large number of parameters in the existing large models. The present invention focuses on the forecasting model in the meteorological field and is also applicable to other fields.

Description

Technical Field

[0001] The present invention belongs to the technical field of meteorological forecasting, and particularly relates to a rapid forecasting method, system, device, medium and terminal of a large meteorological forecasting model with low computing resources. Background Art

[0002] In recent years, large artificial intelligence models have demonstrated astonishing levels of intelligence. However, the current training of large artificial models requires long-term and sufficient computing resource support. The allocation of a large amount of computing resources has greatly hindered a large number of ordinary research institutions from conducting research on large models. Secondly, in terms of model structure, the basic architecture of these large models is the transformer, and the model architecture is single; in terms of training difficulty, for these large models, a large amount of data and a huge number of parameters mean that the training difficulty of the models is extremely high. In addition, the training of these large models also faces complex engineering problems, such as model training under a distributed framework, how to split the model and data, and how to effectively communicate between different nodes. Solving these problems requires the close cooperation of professionals from multiple industries, which further raises the industry barriers of large models.

[0003] Through the above analysis, the problems and defects existing in the prior art are as follows:

[0004] (1) Existing large models are all integrated large model architectures, which have a huge number of parameters, require a large amount of data for training, and have a high training difficulty.

[0005] (2) Existing large model training has high requirements for computing resources.

[0006] (3) Data requires a large amount of manual participation in annotation, and the model training needs to be fine-tuned in combination with expert experience and reinforcement learning technology during the training process.

[0007] (4) Model training requires the participation of professionals in multiple fields, involving the close cooperation of professionals in multiple professional fields such as artificial intelligence, communication maintenance, and computing resource management. Summary of the Invention

[0008] Meteorology is closely related to human activities. Accurate meteorological forecasting is of great significance to fields such as daily travel and urban planning. Current meteorological forecasting is mainly numerical forecasting or AI forecasting, and there are a series of problems such as short forecasting time, low prediction accuracy, and strong model dependence. The powerful potential demonstrated by large artificial intelligence models provides a good solution for improving meteorological forecasting capabilities.

[0009] Aiming at the problems existing in the prior art, the present invention provides a rapid forecasting method, system, device, medium and terminal of a large meteorological forecasting model with low computing resources.

[0010] The present invention is implemented as follows. A fast weather forecasting method for a large weather forecasting model with low computing resources includes:

[0011] Step 1: Construct a dataset for the large weather forecasting model;

[0012] Step 2: Split the large weather forecasting model task into subtasks 1, 2, …, N, and determine the training sets for subtasks 1 - N;

[0013] Step 3: Decompose the output of each subtask, and determine the subtask output branches 1, 2, …, Q for each subtask;

[0014] Step 4: Collect multiple small models that can perform the same task for each subtask type, and construct a small model alternative set;

[0015] Step 5: Select subtask type N, and initialize N to 1;

[0016] Step 6: For the subtask N selected in Step 5, train each small model in the small model alternative set for the subtask output branch 1 of this subtask, and determine the small model with the best performance as the base model;

[0017] Step 7: Fine - tune the small models on the remaining output branches of subtask N: On the base model in Step 6, the remaining output branches of subtask N adopt the same small model structure as the base model, initialize the parameters of the small models on the remaining output branches of subtask N with the parameters of the base model trained in Step 6, and perform parameter fine - tuning;

[0018] Step 8: Construct an integrated small model for subtask N: Based on the small models trained on all output branches of subtask N in Step 6 and Step 7, use the same small model structure to construct an integrated small model for subtask N, and input the outputs of all output branches under subtask N into the integrated small model for subtask N; Initialize the integrated small model for subtask N with the parameters of the base model in Step 6, and perform parameter fine - tuning;

[0019] Step 9: Update subtask type N = N + 1, and repeat Steps 6 - 8 to train the integrated module for the current subtask N;

[0020] Step 10: Integrate the output of the integrated modules of subtask N and N - 1, select the same network structure as the base model in Step 6, initialize the integrated small models of subtask N and N - 1 with the parameters of the base model in Step 6, and fine - tune the parameters of the integrated small models of subtask N and N - 1;

[0021] Step 11: Repeat Steps 9 - 10 until all subtasks are integrated.

[0022] Furthermore, Step 1 includes the following steps:

[0023] S11. Determine the form of the training data according to the task type, collect a large amount of data in the determined form, and form an initial data set.

[0024] S12. Preprocess the initial data set collected in S11, such as screening, enhancement, denoising, annotation, etc., to form a preprocessed data set.

[0025] S13. Divide the preprocessed data set in S12 into a large model training set and a large model test set at a certain ratio.

[0026] Furthermore, step 2 includes the following steps:

[0027] S21. Analyze the large model task and determine the output form of the large model.

[0028] S22. Evenly split the output of the large model in S21 into multiple sub-outputs, and regard each sub-output as a sub-task.

[0029] S23. Output the sub-task types, sub-task 1, sub-task 2,..., sub-task N.

[0030] S24. Construct a sub-task training set. For sub-tasks 1 - N, a certain proportion of random sampling is performed on each sub-task from the large model training set determined in S13 of step 1 to form the training set for each sub-task.

[0031] Furthermore, step 3 includes the following steps:

[0032] S31. Based on the sub-tasks output in S23 of step 2, evenly split the output of each sub-task.

[0033] S32. Determine the output branches of each sub-task after splitting: sub-task output branch 1, sub-task output branch 2,..., sub-task output branch Q.

[0034] S33. Construct the training sets corresponding to different output branches under each sub-task: According to sub-task output branches 1 - Q in S32, a certain proportion of random sampling is performed on each output branch from the sub-task training sets determined in S24 of step 2, so as to construct the training sets corresponding to different output branches of the sub-task: sub-task training set 1, sub-task training set 2,..., sub-task training set Q.

[0035] Furthermore, step 4 includes the following steps:

[0036] S41. For the sub-task types determined in S22 of step 2, determine the input and output data forms of the model.

[0037] S42. Construct a set of alternative small models. Refer to the model input and output data formats determined in S41 and collect various small models with similar input and output data formats.

[0038] Further, step 5 includes the following steps:

[0039] S51. Select the subtask type N and initialize N to 1.

[0040] Step 6 includes the following steps:

[0041] S61. Select the output branch 1 of subtask N and select a small model from the set of alternative small models.

[0042] S62. Configure the training of the small model. Select the optimizer for the small model and set the learning rate, weight decay coefficient, and adaptive moment estimation parameters of the optimizer.

[0043] S63. Determine the loss function of subtask N according to the specific task format of selected subtask N.

[0044] S64. According to S33 in step 3, select the training set 1 of subtask N corresponding to the output branch 1 of subtask N.

[0045] S65. Input the training set in S64 into the small model of the output branch 1 of subtask N and calculate the training loss with the loss function determined in S63.

[0046] S66. According to the training loss in S65, execute the backpropagation algorithm to obtain the parameter update gradient of the small model of the output branch 1 of subtask N.

[0047] S67. Use the optimizer configured in S62 to update the parameters of the small model of the output branch 1 of subtask N until the loss calculated in S65 converges, and save the small model of the output branch 1 of subtask N after convergence during the training process.

[0048] S68. Repeat steps S61 - S67, select the small model with convergent training, and use this small model as the base model of subtask N.

[0049] Step 7 includes the following steps:

[0050] S71. Sequentially select the output branches q of subtask N, where q = 2,..., Q.

[0051] S72. Select the base model determined in S68 in step 6 as the fine-tuning small model of the output branch q of subtask N.

[0052] S73. Fine-tune the training configuration of the small model of the output branch q of subtask N. Select the optimizer for the small model of the output branch q of subtask N and set the learning rate, weight decay coefficient, and adaptive moment estimation parameters of the optimizer.

[0053] S74. Determine the loss function of the output branch q of subtask N according to the task type of the output branch q of subtask N selected by the selection subtask N.

[0054] S75. According to S33 in step 3, select the training set q of subtask N corresponding to the output branch q of subtask N.

[0055] S76. Initialize the small model of the output branch q of subtask N in S72 with the basic model parameters determined in S68 in step 6.

[0056] S77. Input the training set in S75 into the small model of the output branch q of subtask N, and calculate the training loss with the loss function determined in S74.

[0057] S78. According to the training loss in S77, execute the backpropagation algorithm to obtain the parameter update gradient of the small model on the output branch q of subtask N.

[0058] S79. Use the optimizer configured in S73 to update the parameters of the small model on the output branch q of subtask N until the loss calculated in S77 converges, and save the converged small model during the training process.

[0059] S710. Repeat steps S71 - S79 until the training of the small models of all remaining output branches on subtask N is completed.

[0060] Step 8 includes the following steps:

[0061] S81. Obtain the basic model of subtask N in S68 in step 6 and the small models of all remaining output branches on subtask N in S710 in step 7 respectively.

[0062] S82. Set the integrated small model structure of subtask N as the basic model.

[0063] S83. Select all the training data of subtask N as the training set of the integrated small model of subtask N in S82.

[0064] S84. Select the optimizer of the integrated small model of subtask N, and set the learning rate, weight decay coefficient, and adaptive moment estimation parameters of the optimizer.

[0065] S85. According to the task type of subtask N, select the loss function of the integrated small model of subtask N.

[0066] S86. Initialize the integrated small model of subtask N in S82 with the basic model parameters determined in S68 in step 6.

[0067] S87. Input the training set in S83 into output branch 1-Q of subtask N, and input the output of the small model on output branch 1-Q into the integrated small model of subtask N, and calculate the training loss with the loss function determined in S85;

[0068] S88. According to the training loss in S87, execute the backpropagation algorithm to obtain the parameter update gradient of the integrated small model of subtask N;

[0069] S89. With the optimizer configured in S84, update the parameters of the integrated small model of subtask N until the loss change calculated in S87 is stable, and save the minimum loss during the training process and the corresponding integrated small model of subtask N;

[0070] Step 9 includes the following steps:

[0071] S91. Update the type of subtask N, and set N = N + 1;

[0072] S92. Repeat steps 6-8 until the training of all subtask integration modules is completed;

[0073] Step 10 includes the following steps:

[0074] S1001. Obtain the integration modules of subtask N and N-1;

[0075] S1002. Select the optimizer for the integrated small models of subtask N and N-1, and set the learning rate, weight decay coefficient, and adaptive moment estimation parameters of the optimizer;

[0076] S1003. According to the task types of subtask N and N-1, select the loss function for the integrated small models of subtask N and N-1;

[0077] S1004. Initialize the integrated small models of subtask N and N-1 in S1002 with the basic model parameters determined in S68 of step 6;

[0078] S1005. Input the training set in S1002 into the integration module of subtask N and N-1, input the output of the integration module of subtask N and N-1 into the integrated small models of subtask N and N-1, and calculate the training loss with the loss function determined in S1003;

[0079] S1006. According to the training loss in S1005, execute the backpropagation algorithm to obtain the parameter update gradient of the integrated small models of subtask N and N-1;

[0080] S1007. With the optimizer configured in S1002, update the parameters of the integrated small models of subtask N and N-1 until the loss calculated in S1005 converges, and save the integrated small models of subtask N and N-1 corresponding after the training converges;

[0081] Step 11 includes the following steps:

[0082] S1101, repeat Step 9 to save the integration modules of all remaining subtasks;

[0083] S1102, update the subtask N type, N = N + 1;

[0084] S1103, repeat Step 10 to complete the training of the integration small model for subtasks N and N - 1;

[0085] S1104, repeatedly execute S1102 - S1103 until the training of all subtask integration small models is completed.

[0086] Another object of the present invention is to provide a fast weather forecasting large model method under low computing resources, including:

[0087] A dataset construction module for constructing a weather forecasting large model dataset;

[0088] A splitting module for splitting the weather forecasting large model task into subtask 1, subtask 2,..., subtask N, and determining the training sets for subtasks 1 - N;

[0089] A decomposition module for decomposing the output of each subtask to determine the subtask output branches 1, subtask output branches 2,..., subtask output branches Q for each subtask;

[0090] An alternative set construction module for collecting multiple small models that can perform the same task for the subtask type and constructing a small model alternative set;

[0091] An initialization module for selecting the subtask type N and initializing N to 1;

[0092] A training module for, for the selected subtask N, training each small model in the small model alternative set for the output branch 1 of this subtask to determine the small model with the optimal performance as the base model;

[0093] A fine-tuning module for fine-tuning the small models on the remaining output branches of subtask N: on the base model, the remaining output branches of subtask N adopt the same small model structure as the base model, initialize the small model parameters on the remaining output branches of subtask N with the trained base model parameters, and perform parameter fine-tuning; construct the subtask N integration small model: based on the small models trained on all output branches of subtask N, use the same small model structure for the subtask N integration small model, input the outputs of all output branches under subtask N into the subtask N integration small model; initialize the subtask N integration small model with the base model parameters and perform parameter fine-tuning;

[0094] An update module for updating the subtask type N = N + 1 and training the integration module for the current subtask N.

[0095] An integration module for integrating the output of the integration modules of subtasks N and N - 1, selecting the same network structure as the base model, initializing the integration small models of subtasks N and N - 1 with the parameters of the base model, and fine-tuning the parameters of the integration small models of subtasks N and N - 1; until the integration of all subtasks is completed.

[0096] Another object of the present invention is to provide a computer device, which includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor executes the steps of the method for quickly predicting a large meteorological forecasting model under low computing resources.

[0097] Another object of the present invention is to provide a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps of the method for quickly predicting a large meteorological forecasting model under low computing resources.

[0098] Another object of the present invention is to provide an information data processing terminal for implementing the rapid construction of a large meteorological forecasting model under low computing resources.

[0099] Combined with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by the present invention are as follows:

[0100] First, there are many technical barriers in the construction of current large models, such as the difficulty of training large models with a large number of parameters, the high requirements of large models for data and computing resources, and the need for cooperation of experts in multiple fields. To break these technical barriers of large models, this solution focuses on the meteorological field and proposes a method for quickly predicting a large meteorological forecasting model under low computing resources. Based on the fine-tuning technology of neural networks and using the integration idea, it is possible to quickly construct an effective large meteorological forecasting model from scratch under low computing resources.

[0101] The present invention can realize the rapid construction of a large model under low computing resources, avoiding the high computing resource requirements of existing large model training.

[0102] The present invention solves the problem of the difficulty of training large models with a large number of parameters by integrating multiple small models to form a large model.

[0103] The present invention focuses on the forecasting model in the meteorological field and is also applicable to other fields.

[0104] Second, the expected benefits and commercial values after the transformation of the technical solution of the present invention are as follows:

[0105] Greatly reduce the computational resource requirements for large model training, save training costs, and reduce the consumption of material resources.

[0106] Reduce the training difficulty of large models and reduce the human resource consumption in large model training.

[0107] Lower the industry barriers in large model research, enabling research institutions in special computational industries such as meteorology, manufacturing, and healthcare, which have a large amount of data but limited computational resources, to also develop effective industry large models and empower the intelligent development of the industry.

[0108] The technical solution of the present invention fills the technical gap in the domestic and international industries:

[0109] Existing large model training usually requires a large amount of computational resources, a long training time, and the participation and cooperation of experts in multiple fields. This solution can quickly predict effective large models with low computational resources, filling the technical gap of the inability to quickly predict large models under existing low computational resources.

[0110] The technical solution of the present invention solves the technical problems that people have been eager to solve but have never been successful in:

[0111] The solution of the present invention conforms to the actual situation of most domestic research institutions, realizes the rapid and effective training of large models with low resources, and reduces the industry barriers of large models.

[0112] Third, the existing technical problems solved by the technical solution of the present invention in industrial applications:

[0113] 1. The problem of high computational resource requirements

[0114] Existing large meteorological prediction models usually rely on large-scale computational resources such as high-performance computers and GPU clusters, which are difficult for ordinary institutions or enterprises to afford. The present invention greatly reduces the dependence on computational resources through task decomposition and modular small model construction, enabling efficient model training and inference in low-computing power environments.

[0115] 2. The balance problem between model complexity and performance optimization

[0116] Due to its holistic design, traditional meteorological models are difficult to achieve a balance between performance and model complexity. The present invention simplifies the model structure and improves the overall performance by splitting tasks into multiple subtasks and training small models for each subtask.

[0117] 3. The problem of difficulty in optimizing for specific scenarios

[0118] Existing weather forecasting models are usually general models and are difficult to adapt to specific regional or task requirements. Through the decomposition of subtasks and a diverse set of small model alternatives, the present invention achieves efficient optimization for specific meteorological variables or scenarios, improving the pertinence and adaptability of the forecasts.

[0119] 4. Problem of low efficiency in constructing integrated models

[0120] In traditional methods, the construction and optimization of multi-subtask integrated models usually require multiple independent trainings, resulting in low efficiency. Through fine-tuning and parameter inheritance techniques, the present invention quickly transfers the training results of the base model to the subtask output branches and the integration module, significantly shortening the model training time.

[0121] Significant technological advancements of the present invention:

[0122] 1. Significant improvement in resource utilization efficiency

[0123] Through modular task decomposition and small model selection, the present invention achieves forecasting performance close to that of traditional large-scale weather models under low computing power conditions, significantly improving resource utilization efficiency and reducing the threshold for industrial deployment.

[0124] 2. Improvement in model accuracy and adaptability

[0125] Through the refinement of subtasks and the optimization of branch outputs, each subtask is supported by a specially optimized small model, and finally an overall model is formed through the subtask integration module, greatly improving the accuracy and scenario adaptability of weather forecasting.

[0126] 3. Improvement in training efficiency and scalability

[0127] The present invention adopts base model parameter inheritance and fine-tuning techniques in model construction, reducing the computational amount of repeated training and improving training efficiency. At the same time, the task decomposition method makes the expansion of the model more flexible and can quickly adapt to new variables or tasks.

[0128] 4. Applicable to multiple industrial scenarios

[0129] This technical solution is not only applicable to large-scale weather forecasting systems but can also be extended to small-scale and regional forecasting requirements, such as agricultural meteorological services, urban meteorological disaster warnings, etc. scenarios, and has broad industrial application value.

[0130] 5. Reduction of R & D and deployment costs

[0131] Compared with traditional large-scale weather models that require expensive equipment and long-term development, the present invention greatly reduces the R & D cycle and deployment costs through small model selection and modular training methods, providing feasibility for more enterprises and institutions.

[0132] By solving the key problems of the prior art and achieving significant technological progress, the present invention has important application value and competitive advantages in the meteorological forecasting industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0133] Figure 1 It is a flowchart of a rapid forecasting method for a large meteorological forecasting model with low computing resources provided by an embodiment of the present invention.

[0134] Figure 2 It is a structural block diagram of a rapid forecasting system for a large meteorological forecasting model with low computing resources provided by an embodiment of the present invention.

[0135] Figure 3 It is a schematic diagram of the principle of a rapid forecasting method for a large model with low computing resources provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0136] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0137] As Figure 1 shown, a rapid forecasting method for a large meteorological forecasting model with low computing resources provided by an embodiment of the present invention includes the following steps:

[0138] S101. Construct a dataset for the large meteorological forecasting model;

[0139] S102. Split the large meteorological forecasting model task into subtasks 1, subtasks 2,..., subtasks N, and determine the training sets for subtasks 1 - N;

[0140] S103. Decompose the output of each subtask, and determine the subtask output branches 1, subtask output branches 2,..., subtask output branches Q for each subtask;

[0141] S104. Collect multiple small models that perform the same task for the subtask type, and construct a small model alternative set;

[0142] S105. Select subtask type N, and initialize N to 1;

[0143] S106. For the subtask type N selected in S105, train each small model in the small model alternative set for the subtask output branch 1 of this subtask, and determine the small model with the optimal performance as the basic model;

[0144] S107. Fine-tune the small models on the remaining output branches of subtask N: Based on the base model in S106, the other output branches of subtask N adopt the same small model structure as the base model. Initialize the parameters of the small models on the other output branches of subtask N with the parameters of the base model trained in S106, and perform parameter fine-tuning.

[0145] S108. Construct the integrated small model for subtask N: Based on the small models trained on all output branches of subtask N in S106 and S107, use the same small model structure to construct the integrated small model for subtask N. Input the outputs of all output branches under subtask N into the integrated small model for subtask N. Initialize the parameters of the integrated small model for subtask N with the parameters of the base model in S106, and perform parameter fine-tuning.

[0146] S109. Update the subtask type N = N + 1, repeat S106 - S108, and train the integrated model of the current subtask N.

[0147] S110. Integrate the output of the integrated models of subtask N and N - 1, select the network structure same as the base model in S106, initialize the parameters of the integrated small models of subtask N and N - 1 with the parameters of the base model in S106, and fine-tune the parameters of the integrated small models of subtask N and N - 1.

[0148] S111. Repeat S109 - S110 until all subtasks are integrated.

[0149] The embodiment of the present invention proposes a method for quickly predicting a large meteorological forecast model with low computing resources. By modular task decomposition and small model optimization, it solves the bottleneck in computing resource consumption of traditional meteorological forecast models. The working principle includes four main steps: dataset construction, task decomposition, model selection, parameter optimization, and module integration.

[0150] First, in the dataset construction stage (S101), according to the requirements of the meteorological forecast task, integrate multiple meteorological data sources to form a large-scale dataset covering multiple variables, laying a foundation for subsequent training. Subsequently, decompose the complex meteorological forecast task into multiple independent subtasks (S102), where each subtask predicts for a specific meteorological variable or region, simplifying the overall problem into sub-problems that can be solved independently. The output of each subtask is further decomposed into multiple output branches (S103) to ensure that the model has high adaptability to fine-grained forecast targets.

[0151] Secondly, in the model training stage (S104 - S107), for each subtask, first, collect and evaluate multiple small models, and select the optimal small model as the base model according to the performance (S106). Subsequently, through the fine-tuning mechanism (S107), use the parameters of the base model as the initialization point to further optimize the small models of other output branches of the subtask, so as to achieve efficient training and improve the model performance. This training method of modular small models reduces the computational complexity while ensuring high precision for each subtask model.

[0152] Thirdly, in the model integration stage (S108), integrate all output branches of the subtask into a subtask-level integrated small model through the same small model structure, and perform parameter fine-tuning on the integrated small model to ensure the consistency and cooperation of the internal outputs of the subtask. In this way, each subtask independently constructs an integration module, providing an extensible basic module for the construction of the final large model. Subsequently, the integration between subtasks is carried out step by step in a progressive manner (S109 - S111). The output of the integration modules of subtask N and N - 1 is further integrated through a consistent small model structure, and the fine-tuning mechanism is used to optimize the parameters to achieve efficient cooperation between subtasks.

[0153] Finally, the whole process takes modular integration as the core until the integration modules of all subtasks are constructed. Through distributed subtask processing, optimized model selection, progressive strategies of step-by-step integration, and fine-tuning techniques, the present invention achieves the goal of constructing an efficient meteorological forecasting large model in a low-computing-resource environment, not only significantly reducing resource consumption but also ensuring the accuracy and robustness of the model.

[0154] S101 provided by the embodiment of the present invention includes the following steps:

[0155] S11, according to the task type, determine the form of training data, collect a large amount of data in the determined form, and form an initial data set;

[0156] S12, preprocess the initial data set collected in S11, such as screening, enhancement, denoising, annotation, etc., to form a preprocessed data set;

[0157] S13, divide the preprocessed data set in S12 into a large model training set and a large model test set at a certain ratio.

[0158] S102 provided by the embodiment of the present invention includes the following steps:

[0159] S21, analyze the large model task and determine the output form of the large model;

[0160] S22, evenly split the output of the large model in S21 into multiple sub-outputs, and regard each sub-output as a subtask;

[0161] S23. Output subtask types, subtask 1, subtask 2, …, subtask N;

[0162] S24. Construct a subtask training set. For each of subtasks 1 - N, randomly sample a certain proportion from the large model training set determined in S13 of S101 to form the training set for each subtask;

[0163] Step S103 provided by the embodiment of the present invention includes the following steps:

[0164] S31. Based on the subtasks output in S23 of S102, evenly split the output of each subtask;

[0165] S32. Determine the output branches of each split subtask: subtask output branch 1, subtask output branch 2, …, subtask output branch Q;

[0166] S33. Construct the training sets corresponding to different output branches under each subtask: According to subtask output branches 1 - Q in S32, randomly sample a certain proportion from the subtask training set determined in S24 of S102 for each output branch, so as to construct the training sets corresponding to different output branches of the subtask: subtask training set 1, subtask training set 2, …, subtask training set Q;

[0167] Step S104 provided by the embodiment of the present invention includes the following steps:

[0168] S41. For the subtask types determined in S22 of S102, determine the input and output data forms of the model;

[0169] S42. Construct a candidate set of small models. Refer to the input and output data forms of the model determined in S41, and collect a variety of small models with similar input and output data forms;

[0170] Step S105 provided by the embodiment of the present invention includes the following steps:

[0171] S51. Select subtask type N, and initialize N to 1;

[0172] S106 includes the following steps:

[0173] S61. Select subtask N output branch 1, and select a small model from the candidate set of small models;

[0174] S62. Configure the training of the small model. Select the optimizer of the small model, and set the learning rate, weight decay coefficient, and adaptive moment estimation parameters of the optimizer;

[0175] S63. According to the specific task form of selected subtask N, determine the loss function of subtask N;

[0176] S64. According to S33 in S103, select the subtask N training set 1 corresponding to the output branch 1 of subtask N;

[0177] S65. Input the training set in S64 into the small model of the output branch 1 of subtask N, and calculate the training loss with the loss function determined in S63;

[0178] S66. According to the training loss in S65, execute the backpropagation algorithm to obtain the parameter update gradient of the small model of the output branch 1 of subtask N;

[0179] S67. Use the optimizer configured in S62 to update the parameters of the small model of the output branch 1 of subtask N until the loss change calculated in S65 is stable, and save the minimum loss during the training process and the corresponding small model of the output branch 1 of subtask N;

[0180] S68. Repeat steps S61 - S67, select the small model with the minimum loss, and use this small model as the base model of subtask N;

[0181] S107 includes the following steps:

[0182] S71. Sequentially select the output q branches of subtask N, where q = 2,..., Q;

[0183] S72. Select the base model determined in S68 of S106 as the fine-tuning small model of the output branch q of subtask N;

[0184] S73. Fine-tune the training configuration of the small model of the output branch q of subtask N, select the optimizer of the small model of the output branch q of subtask N, and set the learning rate, weight decay coefficient, and adaptive moment estimation parameters of the optimizer;

[0185] S74. According to the task type of the selected output branch q of subtask N, determine the loss function of the output branch q of subtask N;

[0186] S75. According to S33 in S103, select the subtask N training set q corresponding to the output branch q of subtask N;

[0187] S76. Initialize the small model of the output branch q of subtask N in S72 with the base model parameters determined in S68 of S106;

[0188] S77. Input the training set in S75 into the small model of the output branch q of subtask N, and calculate the training loss with the loss function determined in S74;

[0189] S78. According to the training loss in S77, execute the backpropagation algorithm to obtain the parameter update gradient of the small model on the output branch q of subtask N;

[0190] S79. Update the small model parameters on the output branch q of subtask N using the optimizer configured in S73 until the loss change calculated in S77 becomes stable, and save the minimum loss during the training process and the corresponding small model under the minimum loss.

[0191] S710. Repeat steps S71 - S79 until the training of the small models on all remaining output branches of subtask N is completed.

[0192] S108 includes the following steps:

[0193] S81. Obtain the base model of subtask N in S68 of S106 and the small models on all remaining output branches of subtask N in S710 of S107 respectively.

[0194] S82. Set the integrated small model structure of subtask N as the base model.

[0195] S83. Select all the training data of subtask N as the training set for the integrated small model of subtask N in S82.

[0196] S84. Select the optimizer for the integrated small model of subtask N and set the learning rate, weight decay coefficient, and adaptive moment estimation parameters of the optimizer.

[0197] S85. Select the loss function for the integrated small model of subtask N according to the task type of subtask N.

[0198] S86. Initialize the integrated small model of subtask N in S82 with the base model parameters determined in S68 of S106.

[0199] S87. Input the training set in S83 into output branches 1 - Q of subtask N, input the outputs of the small models on output branches 1 - Q into the integrated small model of subtask N, and calculate the training loss using the loss function determined in S85.

[0200] S88. According to the training loss in S87, execute the backpropagation algorithm to obtain the parameter update gradient of the integrated small model of subtask N.

[0201] S89. Update the parameters of the integrated small model of subtask N using the optimizer configured in S84 until the loss change calculated in S87 becomes stable, and save the minimum loss during the training process and the corresponding integrated small model of subtask N.

[0202] S109 includes the following steps:

[0203] S91. Update the type of subtask N, set N = N + 1.

[0204] S92. Repeat S106 - S108 until the training of the integrated modules for all subtasks is completed.

[0205] S110 includes the following steps:

[0206] S1001, obtain the integration modules of subtasks N and N-1;

[0207] S1002, select the integration small model optimizer for subtasks N and N-1, and set the learning rate, weight decay coefficient, and adaptive moment estimation parameters of the optimizer;

[0208] S1003, according to the task types of subtasks N and N-1, select the integration small model loss function for subtasks N and N-1;

[0209] S1004, initialize the integration small models of subtasks N and N-1 in S1002 with the basic model parameters determined in S106;

[0210] S1005, input the training set in S1002 into the integration module of subtasks N and N-1, and the output of the integration module of subtasks N and N-1 is input into the integration small models of subtasks N and N-1, and calculate the training loss with the loss function determined in S1003;

[0211] S1006, according to the training loss in S1005, execute the backpropagation algorithm to obtain the parameter update gradients of the integration small models of subtasks N and N-1;

[0212] S1007, use the optimizer configured in S1002 to update the parameters of the integration small models of subtasks N and N-1 until the loss calculated in S1005 converges, and save the integration small models corresponding to subtasks N and N-1 after the training converges;

[0213] S111 includes the following steps:

[0214] S1101, repeat S109 and save the integration modules of all the remaining subtasks;

[0215] S1102, update the type of subtask N, N = N + 1;

[0216] S1103, repeat S110 to complete the training of the integration small models of subtasks N and N-1;

[0217] S1104, repeatedly execute S1102 - S1103 until the training of all the integration small models is completed.

[0218] As Figure 2 shown, a fast weather forecasting large model system with low computing resources provided by an embodiment of the present invention includes:

[0219] A dataset construction module, used to construct a weather forecasting large model dataset;

[0220] A splitting module, used to split the meteorological forecast large model task into subtasks 1, subtasks 2, …, subtasks N, and determine the training sets of subtasks 1 - N;

[0221] A decomposition module, used to decompose the output of each subtask, and determine subtask output branches 1, subtask output branches 2, …, subtask output branches Q for each subtask;

[0222] An alternative set construction module, used to collect multiple small models that can perform the same task for the subtask type, and construct a small model alternative set;

[0223] An initialization module, used to select subtask type N and initialize N to 1;

[0224] A training module, used to train each small model in the small model alternative set for the selected subtask N on subtask output branch 1 of this subtask, and determine the small model with the optimal performance as the base model;

[0225] A fine - tuning module, used to fine - tune the small models on the remaining output branches of subtask N: on the base model, the remaining output branches of subtask N adopt the same small model structure as the base model, initialize the parameters of the small models on the remaining output branches of subtask N with the trained base model parameters, and perform parameter fine - tuning; construct the integrated small model for subtask N: based on the small models trained on all output branches of subtask N, use the same small model structure to construct the integrated small model for subtask N, and input the outputs of all output branches under subtask N into the integrated small model for subtask N; initialize the integrated small model for subtask N with the base model parameters and perform parameter fine - tuning;

[0226] An update module, used to update subtask type N = N + 1 and train the current subtask N integration module;

[0227] An integration module, used to integrate the output of the integration modules of subtasks N and N - 1, select the same network structure as the base model, initialize the integrated small models of subtasks N and N - 1 with the base model parameters, and fine - tune the parameters of the integrated small models of subtasks N and N - 1; until all subtask integrations are completed.

[0228] Another object of the present invention is to provide a computer device, the computer device includes a memory and a processor, the memory stores a computer program, when the computer program is executed by the processor, the processor executes the steps of the method for quickly forecasting a large meteorological forecast model under low computing resources.

[0229] Another object of the present invention is to provide a computer - readable storage medium, storing a computer program, when the computer program is executed by a processor, the processor executes the steps of the method for quickly forecasting a large meteorological forecast model under low computing resources.

[0230] Another object of the present invention is to provide an information data processing terminal, which is used to implement the rapid construction of a large meteorological forecasting model under low computing resources.

[0231] Specific implementation of the present invention:

[0232] The technical principle diagram of the present invention is as Figure 3 shown, and the technical solutions adopted are as follows:

[0233] 1. Construct a large meteorological forecasting model dataset;

[0234] 2. Split the large meteorological forecasting model task into subtasks 1, subtasks 2,..., subtasks N, and determine the training sets for subtasks 1-N.

[0235] 3. Decompose the output of each subtask, and determine the subtask output branches 1, subtask output branches 2,..., subtask output branches Q for each subtask.

[0236] 4. Collect multiple small models that can perform the same task for the subtask type, and construct a small model alternative set.

[0237] 5. Select subtask type N, and initialize N to 1.

[0238] 6. For the selected subtask N in step 5, train each small model in the small model alternative set for the subtask output branch 1 of this subtask, and determine the small model with the best performance as the base model.

[0239] 7. Fine-tune the small models on the remaining output branches of subtask N: On the base model in step 6, use the same small model structure as the base model for the other output branches of subtask N, initialize the small model parameters on the other output branches of subtask N with the parameters of the trained base model in step 6, and perform parameter fine-tuning.

[0240] 8. Construct an integrated small model for subtask N: Based on the small models trained on all output branches of subtask N in steps 6 and 7, use the same small model structure to construct an integrated small model for subtask N, and input the outputs of all output branches under subtask N into the integrated small model for subtask N. Initialize the integrated small model for subtask N with the base model parameters in step 6, and perform parameter fine-tuning.

[0241] 9. Update subtask type N = N + 1, and repeat steps 5-8 to train the integrated module for the current subtask N.

[0242] 10. Integrate the output of the integrated modules of subtasks N and N-1, select the same network structure as the base model in step 6, initialize the integrated small models of subtasks N and N-1 with the base model parameters in step 6, and fine-tune the parameters of the integrated small models of subtasks N and N-1.

[0243] 11. Repeat steps 9 - 10 until all subtasks are integrated and completed.

[0244] Figure 3 Schematic diagram of the fast prediction method for large models under low computing resources

[0245] Step 1 includes the following steps:

[0246] S11. According to the task type, determine the form of training data, collect a large amount of data in the determined form, and form an initial data set.

[0247] S12. Preprocess the initial data set collected in S11, such as screening, enhancement, denoising, annotation, etc., to form a preprocessed data set.

[0248] S13. Divide the preprocessed data set in S12 into a large model training set and a large model test set at a certain ratio.

[0249] Step 2 includes the following steps:

[0250] S21. Analyze the large model task and determine the output form of the large model.

[0251] S22. Evenly split the output of the large model in S21 into multiple sub - outputs, and regard each sub - output as a subtask.

[0252] S23. Output the subtask types, subtask 1, subtask 2, …, subtask N.

[0253] S24. Construct a subtask training set. For each of the subtasks 1 - N, randomly sample a certain proportion from the large model training set determined in S13 of Step 1 to form the training set for each subtask.

[0254] Step 3 includes the following steps:

[0255] S31. Based on the subtasks output in S23 of Step 2, evenly split the output of each subtask.

[0256] S32. Determine the output branches of each split subtask: subtask output branch 1, subtask output branch 2, …, subtask output branch Q.

[0257] S33. Construct the training sets corresponding to different output branches under each subtask: According to the subtask output branches 1 - Q in S32, randomly sample a certain proportion from the subtask training sets determined in S24 of Step 2 for each output branch, so as to construct the training sets corresponding to different output branches of the subtask: subtask training set 1, subtask training set 2, …, subtask training set Q.

[0258] Step 4 includes the following steps:

[0259] S41. Determine the input and output data formats of the model according to the sub-task type determined in S22 of step 2.

[0260] S42. Construct a set of alternative small models. Referring to the input and output data formats of the model determined in S41, collect multiple small models with similar input and output data formats.

[0261] Step 5 includes the following steps:

[0262] S51. Select sub-task type N and initialize N to 1.

[0263] Step 6 includes the following steps:

[0264] S61. Select output branch 1 of sub-task N and select a small model from the set of alternative small models.

[0265] S62. Configure the training of the small model. Select the optimizer for the small model and set the learning rate, weight decay coefficient, and adaptive moment estimation parameters of the optimizer.

[0266] S63. Determine the loss function of sub-task N according to the specific task format of selected sub-task N.

[0267] S64. According to S33 in step 3, select training set 1 of sub-task N corresponding to output branch 1 of sub-task N.

[0268] S65. Input the training set in S64 into the small model of output branch 1 of sub-task N and calculate the training loss with the loss function determined in S63.

[0269] S66. According to the training loss in S65, execute the backpropagation algorithm to obtain the parameter update gradient of the small model of output branch 1 of sub-task N.

[0270] S67. Use the optimizer configured in S62 to update the parameters of the small model of output branch 1 of sub-task N until the loss change calculated in S65 is stable, and save the minimum loss during the training process and the corresponding small model of output branch 1 of sub-task N.

[0271] S68. Repeat steps S61 - S67, select the small model with the minimum loss, and use this small model as the base model of sub-task N.

[0272] Step 7 includes the following steps:

[0273] S71. Sequentially select output branches q of sub-task N, where q = 2, …, Q.

[0274] S72. Select the base model determined in S68 of step 6 as the fine-tuning small model for output branch q of sub-task N.

[0275] S73. Fine-tune the training configuration of the small model for the output branch q of subtask N, select the optimizer for the small model of the output branch q of subtask N, and set the learning rate, weight decay coefficient, and adaptive moment estimation parameters of the optimizer.

[0276] S74. Determine the loss function for the output branch q of subtask N according to the task type of the output branch q of subtask N.

[0277] S75. According to S33 in step 3, select the training set q for subtask N corresponding to the output branch q of subtask N.

[0278] S76. Initialize the small model of the output branch q of subtask N in S72 with the basic model parameters determined in S68 in step 6.

[0279] S77. Input the training set in S75 into the small model of the output branch q of subtask N, and calculate the training loss with the loss function determined in S74.

[0280] S78. According to the training loss in S77, execute the backpropagation algorithm to obtain the parameter update gradient of the small model on the output branch q of subtask N.

[0281] S79. Use the optimizer configured in S73 to update the parameters of the small model on the output branch q of subtask N until the loss change calculated in S77 is stable, and save the minimum loss during the training process and the corresponding small model under the minimum loss.

[0282] S710. Repeat steps S71 - S79 until the training of the small models for all remaining output branches on subtask N is completed.

[0283] Step 8 includes the following steps:

[0284] S81. Obtain the basic model of subtask N in S68 in step 6 and the small models of all remaining output branches on subtask N in S710 in step 7 respectively.

[0285] S82. Set the integrated small model structure of subtask N as the basic model.

[0286] S83. Select all the training data of subtask N as the training set for the integrated small model of subtask N in S82.

[0287] S84. Select the optimizer for the integrated small model of subtask N, and set the learning rate, weight decay coefficient, and adaptive moment estimation parameters of the optimizer.

[0288] S85. Select the loss function for the integrated small model of subtask N according to the task type of subtask N.

[0289] S86, initialize the sub-task N integrated small model in S82 with the basic model parameters determined in S68 of step 6.

[0290] S87, input the training set in S83 into output branch 1-Q of sub-task N, and input the output of the small model on output branch 1-Q into the sub-task N integrated small model, and calculate the training loss with the loss function determined in S85.

[0291] S88, according to the training loss in S87, execute the backpropagation algorithm to obtain the parameter update gradient of the sub-task N integrated small model.

[0292] S89, use the optimizer configured in S84 to update the parameters of the sub-task N integrated small model until the loss change calculated in S87 is stable, and save the minimum loss during the training process and the corresponding sub-task N integrated small model.

[0293] Step 9 includes the following steps:

[0294] S91, update the type of sub-task N, and set N = N + 1.

[0295] S92, repeat steps 5-8 until the integration module training of all sub-tasks is completed.

[0296] Step 10 includes the following steps:

[0297] S1001, obtain the integration modules of sub-task N and N-1.

[0298] S1002, select the optimizer for the integrated small models of sub-task N and N-1, and set the learning rate, weight decay coefficient, and adaptive moment estimation parameters of the optimizer.

[0299] S1003, according to the task types of sub-task N and N-1, select the loss function for the integrated small models of sub-task N and N-1.

[0300] S1004, initialize the integrated small models of sub-task N and N-1 in S1002 with the basic model parameters determined in S68 of step 6.

[0301] S1005, input the training set in S1002 into the integration module of sub-task N and N-1, input the output of the integration module of sub-task N and N-1 into the integrated small models of sub-task N and N-1, and calculate the training loss with the loss function determined in S1003.

[0302] S1006, according to the training loss in S1005, execute the backpropagation algorithm to obtain the parameter update gradient of the integrated small models of sub-task N and N-1.

[0303] S1007, Update the integrated small model parameters of subtasks N and N-1 using the optimizer configured in S1002 until the loss change calculated in S1005 is stable, and save the minimum loss during the training process and the corresponding integrated small models of subtasks N and N-1.

[0304] Step 11 includes the following steps:

[0305] S1101, Repeat step 9 and save the integrated modules of all remaining subtasks.

[0306] S1102, Update the subtask N type, N = N + 1.

[0307] S1103, Repeat step 10 to complete the training of the integrated small models of subtasks N and N-1.

[0308] S1104, Repeat the execution of S1102 - S1103 until the training of all integrated small models of subtasks is completed.

[0309] This solution includes:

[0310] 1. Construct a meteorological large model dataset:

[0311] A11, Collect data on different meteorological elements globally from the ERA5 platform for 10 years (2013 - 2023);

[0312] A12, Process the data in A11 into a tensor format of H*W*C*T, where H, W, C, and T represent the height, width, number of meteorological elements, and time of the data respectively;

[0313] A13, Divide the meteorological data in step A12 from the time dimension in a ratio of 7:3 to construct a training set and a test set.

[0314] 2. Meteorological large model construction and training.

[0315] A21, Determine the subtasks of the meteorological forecasting large model and construct a subtask training set.

[0316] 1.1 Determine that the meteorological forecasting large model task is time series modeling, which inputs all meteorological elements at the previous t0 moments and predicts all meteorological elements at the subsequent t1 moments. The input data is in the tensor form of H*W*C*t0, and the output is H*W*C*(t0 + t1).

[0317] 1.2 Evenly split the output H*W*C*(t0 + t1) in A21 in the time dimension into N sub-outputs, and the data format of each subtask output is: H*W*C*((t0 + t1) / N).

[0318] 1.3. Output subtask types, subtask 1, subtask 2, …, subtask N. The input and output tensor sizes of subtasks 1 - N satisfy: Input tensor size: H*W*C*t0, and output tensor size: H*W*C*((t0 + t1) / N).

[0319] 1.4. Construct a subtask training set. For subtasks 1 - N, for each subtask, 70% random sampling is performed from the large model training set determined in 1.3) to form the training set for this subtask.

[0320] A22, Subtask output splitting, construct a training set for the subtask output branch.

[0321] 2.1. Based on the subtasks in 1.3, evenly split the output of each subtask. The input and output tensor shapes of the subtask output branch are: H*W*C*t0, H*W*C*((t0 + t1) / N / Q).

[0322] 2.2. According to the output f shape of the subtask output branch in 2.1, sequentially determine each subtask output branch: subtask output branch 1, subtask output branch 2, …, subtask output branch Q.

[0323] 2.3. According to subtask output branches 1 - Q in 2.2, for each output branch, 70% of the data is randomly sampled from the subtask training set determined in 1) 4 to construct training sets on different output branches of the subtask: subtask training set 1, subtask training set 2, …, subtask training set Q.

[0324] A23, Determine the small model candidate set

[0325] 3.1. For the task type, input, and output tensor format in 1.1, select candidate small models that can perform the same task and have the same input data format.

[0326] 3.2. Determine multiple candidate small models to construct a small model candidate set.

[0327] A24, Basic model training for subtask N

[0328] 4.1. Select subtask type N and initialize N to 1.

[0329] 4.2. Select output branch 1 of subtask N and select a small model from the candidate small model set.

[0330] 4.3. Small model training configuration. Select Adam as the optimizer for the small model, and set the learning rate of the optimizer to 0.0001, the weight decay coefficient to 0.001, and the adaptive moment estimation parameters to 0.9 and 0.99.

[0331] 4.4. Determine that the loss function of the output branch 1 of sub-task N is MSE.

[0332] 4.5. Select the corresponding sub-task training set 1 of the output branch 1 of sub-task N from 2.3.

[0333] 4.6. Input the training set in 4.5 into the small model, and calculate the training loss with the loss function determined in 4.4.

[0334] 4.7. According to the training loss in 4.6, execute the backpropagation algorithm to obtain the parameter update gradient of the small model.

[0335] 4.8. Use the Adam optimizer configured in 4.3 to update the parameters of the small model until the loss calculated in 4.6 converges, and save the small model corresponding to the convergence of training.

[0336] 4.9. Repeat steps 4.2 - 4.8, select the small model with convergent training, and use this small model as the base model.

[0337] A25, Fine-tune the small models of the remaining output branches of sub-task N

[0338] 5.1. Select the output branches q of sub-task N in sequence, where q = 2, …, Q, and initialize q = 2.

[0339] 5.2. Select the base model in 4.9 as the small model on the output branch q of sub-task N.

[0340] 5.3. For the training configuration of the fine-tuned small model of sub-task N branch q, select Adam as the optimizer of the small model, and set the learning rate of the optimizer to 0.0001, the weight decay coefficient to 0.001, and the adaptive moment estimation parameters to 0.9 and 0.99.

[0341] 5.4. Select MSE as the loss function of the output branch q of sub-task N.

[0342] 5.5. Select the corresponding sub-task training set q of the output branch q of sub-task N from 2.3.

[0343] 5.6. Initialize the small model on the output branch q of sub-task N in 5.2 with the parameters of the base model in 4.9.

[0344] 5.7. Input the training set in 5.5 into the small model on the output branch q of sub-task N, and calculate the training loss with the loss function determined in 5.4.

[0345] 5.8. According to the training loss in 5.7, execute the backpropagation algorithm to obtain the update gradient of the parameters of the small model on the output branch q of sub-task N.

[0346] 5.9. Update the parameters of the small model on the output branch q of subtask N with the optimizer configured in 5.3 until the loss change calculated in 5.7 is stable, and save the minimum loss during training and the corresponding small model on the output branch q of subtask N.

[0347] 5.10. Update the output branch q of subtask N, q = q + 1, and repeat steps 5.1 - 5.9 until the small models on all remaining output branches of subtask N are trained.

[0348] A26, Fine-tune the integrated small model of subtask N

[0349] 6.1. Obtain the base model on the output branch 1 of subtask N and the small models on the remaining output branches of subtask N according to 4.9 and 5.10 respectively.

[0350] 6.2. Set the structure of the integrated small model of subtask N as the base model.

[0351] 6.3. Select all the training data of subtask N from 1.4 as the training set of the integrated small model of subtask N in 6.2.

[0352] 6.4. Select Adam as the optimizer for the integrated small model of subtask N, and set the learning rate of the optimizer to 0.0001, the weight decay coefficient to 0.001, and the adaptive moment estimation parameters to 0.9 and 0.99.

[0353] 6.5. Select the MSE as the loss function of the integrated small model of subtask N.

[0354] 6.6. Initialize the integrated small model of subtask N in 6.2 with the parameters of the base model in 4.9.

[0355] 6.7. Input the training set of subtask N in 6.3 into the small models on the output branches 1 - Q of subtask N, and input the outputs of the small models on the output branches 1 - Q of subtask N into the integrated small model of subtask N, and calculate the training loss with the loss function determined in 6.5.

[0356] 6.8. According to the training loss in 6.7, execute the backpropagation algorithm to obtain the parameter update gradient of the integrated small model of subtask N.

[0357] 6.9. Update the parameters of the integrated small model of subtask N with the optimizer configured in 6.4 until the loss change calculated in 6.7 is stable, and save the minimum loss during the training process and the corresponding integrated small model of subtask N.

[0358] A27, Train the integrated modules of other subtasks

[0359] 7.1. Update the type of subtask N, set N = N + 1.

[0360] 7.2. Repeat steps A22 - A26 until all remaining sub - task integration modules are completed.

[0361] A28, Integration of sub - tasks N and N - 1

[0362] 8.1. Obtain the integration modules for sub - tasks N and N - 1.

[0363] 8.2. Set the integration small - model structure on sub - tasks N and N - 1 to the basic small - model.

[0364] 8.3. Select all training data on sub - tasks N and N - 1 from 1.4 as the training set for the integration small - model of sub - tasks N and N - 1 in 8.2.

[0365] 8.4. Set the optimizer for the integration small - model of sub - tasks N and N - 1 in 8.2 to Adam, and set the learning rate of the optimizer to 0.0001, the weight decay coefficient to 0.001, and the adaptive moment estimation parameters to 0.9 and 0.99.

[0366] 8.5. Select MSE as the loss function for the integration small - model on sub - tasks N and N - 1.

[0367] 8.6. Initialize the integration small - model of sub - tasks N and N - 1 in 8.2 with the basic model parameters in 4.9.

[0368] 8.7. Input the training set in 8.3 into the integration module of sub - tasks N and N - 1, input the output of the integration module of sub - tasks N and N - 1 into the integration small - model of sub - tasks N and N - 1, and calculate the training loss with the loss function determined in 8.5.

[0369] 8.8. According to the training loss in 8.7, execute the backpropagation algorithm to obtain the parameter update gradients of the integration small - model of sub - tasks N and N - 1.

[0370] 8.9. With the optimizer configured in 8.4, update the parameters of the integration small - model of sub - tasks N and N - 1 until the change in the loss calculated in 8.7 is stable, and save the minimum loss during the training process and the corresponding integration small - model of sub - tasks N and N - 1.

[0371] A29, Integration of remaining sub - tasks

[0372] 9.1. Repeat A27 to complete the training of the integration modules for all remaining sub - tasks.

[0373] 9.2. Update the type of sub - task N, N = N + 1.

[0374] 9.3. Repeat A28 to complete the training of the integration small - model for sub - tasks N and N - 1.

[0375] 9.4. Repeat the execution of 9.2 - 9.3 until the training of all sub - task integrated small models is completed.

[0376] 3. Testing of the large - scale meteorological forecasting model

[0377] A31, Input the test data into Sub - task 1, Sub - task 2, …, Sub - task N in sequence.

[0378] A32, Obtain the output results of the integrated small models of Sub - task 1, 2, …, N.

[0379] The present invention provides a fast forecasting method for large - scale models with low computing resources, which successfully overcomes the problem of high computing resource requirements in the training process of existing large - scale models. This method avoids the bottleneck of traditional large - scale models in terms of hardware resources and provides the possibility for developing large - scale meteorological forecasting models in scenarios with limited computing resources. Secondly, the present invention constructs a large - scale model by integrating multiple small models, significantly reducing the training difficulty brought by a large number of parameters. By decomposing complex problems into sub - task modules and training and integrating them step by step, the efficiency and scalability of model training are improved. In addition, although the present invention focuses on the forecasting model in the meteorological field, its methodology is also applicable to the construction of large - scale models in other fields, with strong generality and cross - field adaptation ability.

[0380] The technical solution in the present invention has a certain degree of flexibility, and the following alternative solutions can be specifically used to achieve similar invention purposes:

[0381] 1. Replace the basic model: This solution uses the TAU model as the basic small model. However, for different tasks, it can be flexibly replaced with other lightweight small models, such as network structures like LSTM, Transformer, or CNN, and adjust the model structure and configuration according to specific task requirements.

[0382] 2. Retrain the sub - task models: In the current solution, the training of each sub - task branch small model and sub - task integrated model adopts the fine - tuning training method. In the alternative solution, a retraining method from scratch can be adopted. Although it requires a longer training time, it can perform a more refined model fitting for new data.

[0383] 3. Optimize the training configuration: During the training process of the basic model, sub - task branch models, and integrated models, other types of optimizers (such as AdamW, SGD, etc.) can be selected, and the learning rate or weight decay parameters can be adjusted to improve the convergence speed and prediction accuracy of the model.

[0384] The technical key points of the present invention lie in its large - scale model integration design method with low computing resources and the integrated training method suitable for a large number of parameters. It mainly includes the following points:

[0385] 1. Large model integrated design method: By decomposing the prediction task of future meteorological variables into subtask modules and subtask branches, the problem of computational resource limitations caused by the high number of parameters in large models is solved. Each subtask branch model focuses on the prediction of specific time periods and variable dimensions, significantly improving the training efficiency and model performance.

[0386] 2. Integrated training method: Adopting a modular design, by independently training subtask models and integrating subtask outputs in the integration stage, the decoupling between subtasks is ensured, reducing the training difficulty of complex models, and at the same time achieving efficient prediction performance.

[0387] Based on ERA5 meteorological data, the present invention proposes a large model construction method, using the TAU model as the basic small model, inputting 20 variables (6 time moments, 4 pressure levels), and outputting 20 meteorological elements corresponding to the future 4, 8, 12, and 16 time moments. The specific model design includes:

[0388] 1. Subtask decomposition: The variable output for the future 16 time moments is decomposed into 4 subtasks, and each subtask is responsible for predicting meteorological variables at 1 - 4, 4 - 8, 8 - 12, and 12 - 16 time moments respectively.

[0389] 2. Subtask branch design: Each subtask is further refined into 4 branches, responsible for outputting meteorological elements at positions 1 - 5, 5 - 10, 10 - 15, and 15 - 20 respectively, forming a finer - grained task division.

[0390] 3. Subtask integration: The outputs of each subtask are gradually integrated into the complete 20 - variable output for the future 1 - 4, 1 - 8, 1 - 12, and 1 - 16 time moments, realizing the collaborative integration between tasks and ensuring the consistency and integrity of the model output.

[0391] Through this design, the present invention decomposes complex meteorological forecasting tasks into multiple subtasks in a modular manner for independent training and collaborative integration, greatly improving the training efficiency and prediction accuracy, being applicable to resource - constrained scenarios, and at the same time having strong scalability and application value.

[0392] Evidence related to the technical effects obtained in the embodiments of the present invention.

[0393] The experimental results are as follows:

[0394] MSE test results of the meteorological forecasting large model

[0395]

[0396] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated designed hardware. Those of ordinary skill in the art can understand that the above devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code is provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and their modules of the present invention can be implemented by hardware circuits of programmable hardware devices such as very large scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, or can be implemented by a combination of the above hardware circuits and software such as firmware.

[0397] As described above, the above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A rapid forecasting method for a large meteorological forecasting model with low computing resources, characterized in that, It includes the following steps: Step 1: Construct a large meteorological forecast model dataset; Step 2: Split the large meteorological forecast model task into subtasks 1, subtasks 2,..., subtasks N, and determine the training sets for subtasks 1 - N; Step 3: Decompose the output of each subtask, and determine the subtask output branches 1, subtask output branches 2,..., subtask output branches Q for each subtask; Step 4: Collect multiple small models that perform the same task for the subtask type, and construct a small model alternative set; Step 5: Select subtask type N, and initialize N to 1; Step 6: For the subtask type N selected in Step 5, train each small model in the small model alternative set for the subtask output branch 1 of this subtask, and determine the small model with the optimal performance as the base model; Step 7: Fine - tune the small models on the remaining output branches of subtask N: On the base model in Step 6, the remaining output branches of subtask N adopt the same small model structure as the base model, initialize the parameters of the small models on the remaining output branches of subtask N with the parameters of the base model trained in Step 6, and perform parameter fine - tuning; Step 8: Construct an integrated small model for subtask N: Based on the small models trained on all output branches of subtask N in Step 6 and Step 7, use the same small model structure to construct an integrated small model for subtask N, and input the outputs of all output branches under subtask N into the integrated small model for subtask N; Initialize the integrated small model for subtask N with the parameters of the base model in Step 6, and perform parameter fine - tuning; Step 9: Update subtask type N = N + 1, repeat Steps 5 - 8 to train the integrated model of the current subtask N; Step 10: Integrate the output of the integrated models of subtasks N and N - 1, select the same network structure as the base model in Step 6, initialize the integrated small models of subtasks N and N - 1 with the parameters of the base model in Step 6, and fine - tune the parameters of the integrated small models of subtasks N and N - 1; Step 11: Repeat Steps 9 - 10 until all subtasks are integrated; Step 5 includes the following steps: S51, Select subtask type N, and initialize N to 1; Step 6 includes the following steps: S61, Select subtask N output branch 1, and select a small model from the alternative small model set; S62, Configure the training of the small model, select the optimizer of the small model, and set the learning rate, weight decay coefficient, and adaptive moment estimation parameters of the optimizer; S63, Determine the loss function of subtask N according to the specific task form of the selected subtask N; S64, According to S33 in Step 3, select the training set 1 of subtask N corresponding to subtask N output branch 1; S65, Input the training set in S64 into the small model of subtask N output branch 1, and calculate the training loss with the loss function determined in S63; S66, According to the training loss in S65, execute the backpropagation algorithm to obtain the parameter update gradient of the small model of subtask N output branch 1; S67, Use the optimizer configured in S62 to update the parameters of the small model of subtask N output branch 1 until the loss change calculated in S65 is stable, save the minimum loss during the training process and the corresponding small model of subtask N output branch 1; S68. Repeat steps S61 - S67, select the small model with the minimum loss, and use this small model as the base model for subtask N; Step 7 includes the following steps: S71. Sequentially select the output q branches of subtask N, where q = 2,..., Q; S72. Select the base model determined in S68 of step 6 as the fine-tuning small model for the output branch q of subtask N; S73. Fine-tune the training configuration of the small model for the output branch q of subtask N, select the optimizer for the small model of the output branch q of subtask N, and set the learning rate, weight decay coefficient, and adaptive moment estimation parameters of the optimizer; S74. Determine the loss function for the output branch q of subtask N according to the task type of the selected output branch q of subtask N; S75. According to S33 in step 3, select the training set q of subtask N corresponding to the output branch q of subtask N; S76. Initialize the small model for the output branch q of subtask N in S72 with the base model parameters determined in S68 of step 6; S77. Input the training set in S75 into the small model for the output branch q of subtask N, and calculate the training loss with the loss function determined in S74; S78. According to the training loss in S77, execute the backpropagation algorithm to obtain the parameter update gradient of the small model on the output branch q of subtask N; S79. Use the optimizer configured in S73 to update the parameters of the small model on the output branch q of subtask N until the loss change calculated in S77 is stable, and save the minimum loss during the training process and the corresponding small model under the minimum loss; S710. Repeat steps S71 - S79 until the training of the small models for all remaining output branches on subtask N is completed; Step 8 includes the following steps: S81. Obtain the base model of subtask N in S68 of step 6 and the small models for all remaining output branches on subtask N in S710 of step 7 respectively; S82. Set the structure of the ensemble small model of subtask N as the base model; S83. Select all the training data of subtask N as the training set for the ensemble small model of subtask N in S82; S84. Select the optimizer for the ensemble small model of subtask N and set the learning rate, weight decay coefficient, and adaptive moment estimation parameters of the optimizer; S85. Select the loss function for the ensemble small model of subtask N according to the task type of subtask N; S86. Initialize the ensemble small model of subtask N in S82 with the base model parameters determined in S68 of step 6; S87. Input the training set in S83 into the output branches 1 - Q of subtask N, input the outputs of the small models on the output branches 1 - Q into the ensemble small model of subtask N, and calculate the training loss with the loss function determined in S85; S88. According to the training loss in S87, execute the backpropagation algorithm to obtain the parameter update gradient of the ensemble small model of subtask N; S89. Use the optimizer configured in S84 to update the parameters of the ensemble small model of subtask N until the loss change calculated in S87 is stable, and save the minimum loss during the training process and the corresponding ensemble small model of subtask N; Step 9 includes the following steps: S91. Update the type of subtask N, set N = N + 1; S92. Repeat steps 5 - 8 until the integrated module training for all subtasks is completed; Step 10 includes the following steps: S1001. Obtain the integrated modules for subtasks N and N - 1; S1002. Select the optimizer for the integrated small models of subtasks N and N - 1, and set the learning rate, weight decay coefficient, and adaptive moment estimation parameters of the optimizer; S1003. Select the loss function for the integrated small models of subtasks N and N - 1 according to the task types of subtasks N and N - 1; S1004. Initialize the integrated small models of subtasks N and N - 1 in S1002 with the basic model parameters determined in S68 of step 6; S1005. Input the training set in S1002 into the integrated module of subtasks N and N - 1, and input the output of the integrated module of subtasks N and N - 1 into the integrated small models of subtasks N and N - 1, and calculate the training loss with the loss function determined in S1003; S1006. According to the training loss in S1005, execute the backpropagation algorithm to obtain the parameter update gradients of the integrated small models of subtasks N and N - 1; S1007. Use the optimizer configured in S1002 to update the parameters of the integrated small models of subtasks N and N - 1 until the change in the loss calculated in S1005 is stable, and save the minimum loss during the training process and the corresponding integrated small models of subtasks N and N - 1; Step 11 includes the following steps: S1101. Repeat step 9 and save the integrated modules of all remaining subtasks; S1102. Update the subtask N type, N = N + 1; S1103. Repeat step 10 to complete the training of the integrated small models of subtasks N and N - 1; S1104. Repeat the execution of S1102 - S1103 until the training of all integrated small models of subtasks is completed.

2. The rapid weather forecasting method for large meteorological forecasting models under low computing resources according to claim 1, characterized in that The said step 1 includes the following steps: S11. According to the subtask types, collect and determine the formal data to form an initial data set; S12. Preprocess the initial data set collected in S11, including screening, enhancement, denoising, and annotation, to form a preprocessed data set; S13. Divide the preprocessed data set in S12 into a training set and a test set for the weather forecasting large model.

3. The rapid weather forecasting method for large meteorological forecasting models under low computing resources according to claim 1, characterized in that The said step 2 includes the following steps: S21. Analyze the large model task and determine the output form of the large model; S22. Evenly split the output of the large model in S21 into multiple sub - outputs, and regard each sub - output as a subtask; S23. Output the subtask types, subtask 1, subtask 2,..., subtask N; Decompose the output of each subtask to determine the subtask output branches 1, 2,..., Q of each subtask; S24. Construct a subtask training set. For subtasks 1 - N, randomly sample from the training set of the large model determined in S13 of step 1 for each subtask to form the training set of this subtask.

4. The rapid weather forecasting method for large meteorological forecasting models under low computing resources according to claim 1, characterized in that, The said step 3 includes the following steps: S31. Based on the subtasks output in S23 of step 2, evenly split the output of each subtask; S32. Determine the output branches of each split subtask: subtask output branch 1, subtask output branch 2,..., subtask output branch Q; S33. Construct the training set corresponding to the output branch of each sub-task after splitting: According to the sub-task output branches 1-Q in S32, each output branch randomly samples from the sub-task training set determined in S24 of step 2 to construct the training sets corresponding to the output branches of each sub-task after splitting: sub-task training set 1, sub-task training set 2,..., sub-task training set Q.

5. The rapid weather forecasting method for large weather forecasting models with low computing resources according to claim 1, characterized in that, Step 4 includes the following steps: S41. For the sub-task types determined in S22 of step 2, determine the input and output data forms of the model. S42. Construct a small model alternative set. Referring to the model input and output data forms determined in S41, collect small models with various input and output data forms.

6. A rapid weather forecasting large model system with low computing resources for implementing the rapid weather forecasting method of the large model under low computing resources according to any one of claims 1-5, characterized in that, The large model rapid prediction system for weather forecasting under low computing resources includes: A data set construction module for constructing a large model data set for weather forecasting. A splitting module for splitting the large model task of weather forecasting into sub-tasks 1, sub-tasks 2,..., sub-tasks N, and determining the training sets of sub-tasks 1-N. A decomposition module for decomposing the output of each sub-task to determine the sub-task output branches 1, sub-task output branches 2,..., sub-task output branches Q of each sub-task. An alternative set construction module for collecting multiple small models performing the same task for the sub-task type and constructing a small model alternative set. An initialization module for selecting the sub-task type N and initializing N to 1. A training module for, for the selected sub-task N, training each small model in the small model alternative set for the output branch 1 of this sub-task, and determining the small model with the optimal performance as the base model. A fine-tuning module for fine-tuning the small models on the remaining output branches of sub-task N: On the base model, the remaining output branches of sub-task N adopt the same small model structure as the base model, initialize the parameters of the small models on the remaining output branches of sub-task N with the trained base model parameters, and perform parameter fine-tuning; construct the integrated small model of sub-task N: Based on the small models trained on all output branches of sub-task N, use the same small model structure to integrate the small model for sub-task N, and input the outputs of all output branches under sub-task N into the integrated small model of sub-task N; initialize the integrated small model of sub-task N with the base model parameters and perform parameter fine-tuning. An update module for updating the sub-task type N = N + 1 and training the integrated module of the current sub-task N. An integration module for integrating the output of the integrated modules of sub-tasks N and N-1, selecting the same network structure as the base model, initializing the integrated small models of sub-tasks N and N-1 with the base model parameters, and fine-tuning the parameters of the integrated small models of sub-tasks N and N-1; until all sub-tasks are integrated.

7. A computer device, characterized in that, The computer device includes a memory and a processor. When the computer program stored in the memory is executed by the processor, the processor executes the steps of the large model rapid prediction method for weather forecasting under low computing resources as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, A computer program is stored, and when the computer program is executed by a processor, the processor is caused to execute the steps of the method for quickly predicting a large meteorological prediction model under low computing resources as described in any one of claims 1-5.

9. An information data processing terminal, characterized in that, The information data processing terminal is used to implement the system for quickly predicting a large meteorological prediction model under low computing resources as described in claim 6.

Images