Short-term forecast evaluation system and parallel computing optimization method
By combining a short-term forecast assessment system with parallel computing optimization methods, the problem of the disconnect between assessment and computation was solved, enabling dynamic resource allocation and self-optimization, thereby improving the accuracy and efficiency of short-term forecasts.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHENGDU RUNLIAN TECH DEV
- Filing Date
- 2026-02-02
- Publication Date
- 2026-04-17
AI Technical Summary
In existing technologies, the evaluation and computational optimization stages of short-term forecasting systems are disconnected, resulting in inaccurate allocation of computational resources, a lack of adaptive optimization capabilities, and an impact on the overall performance improvement of the forecasting system.
A short-term forecast assessment system is adopted, including a data interface module, a dynamic assessment module, a resource decision module, and a parallel computing cluster. Through multi-dimensional assessment indicators and a policy mapping rule base, dynamic resource allocation and task scheduling are realized, and a closed-loop feedback mechanism of assessment-optimization is constructed.
It enables precise and dynamic allocation of computing resources, enhances the self-optimization capability of the forecasting system, continuously improves forecast accuracy and efficiency, and forms a virtuous cycle.
Smart Images
Figure CN121614276B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of short-term forecast assessment technology, and more specifically, to a short-term forecast assessment system and a parallel computing optimization method. Background Technology
[0002] Short-term forecasting is extremely important in many fields of modern society, such as meteorology, transportation, and power dispatching. Its core is to accurately predict high-impact events in the near future using the latest observational data and computational models within a limited timeframe.
[0003] Currently, with the increasing complexity of forecast models and the explosive growth of data volume, the demand for backend computing power is also rising exponentially. The mainstream solutions in existing technologies mainly rely on high-performance parallel computing clusters. However, existing technical systems typically treat "forecast evaluation" and "computational optimization" as two separate processes: the evaluation system is responsible for verifying forecast accuracy afterward, while computing resources are mostly allocated statically or empirically, resulting in a disconnected processing approach.
[0004] This leads to a disconnect between the allocation of computing resources and the actual performance of forecasts. For example, insufficient computing resources are allocated to critical areas or periods with high forecast uncertainty, while resources are wasted in areas where forecasts are relatively stable. Furthermore, existing forecast systems often lack a closed-loop feedback mechanism for self-optimization based on real-time performance. This results in evaluations being used only for judgment and failing to directly and automatically guide the computing strategy for the next cycle, thus limiting the continuous improvement of the overall performance of the forecast system.
[0005] Therefore, this application proposes a short-term forecast evaluation system and a parallel computing optimization method to solve the above-mentioned technical problems. Summary of the Invention
[0006] The summary section of this invention provides a brief overview of the concepts, which will be described in detail in the detailed description section that follows. This summary section is not intended to identify key or essential features of the claimed invention, nor is it intended to limit the scope of the claimed invention.
[0007] The purpose of this invention is to address the shortcomings of existing technologies by providing a short-term forecast assessment system and a parallel computing optimization method, thereby solving the problems of inaccurate allocation of computing resources and the lack of overall adaptive optimization capabilities of the system caused by the disconnect between the assessment and computation optimization stages.
[0008] To achieve the above objectives, the present invention provides the following technical solution:
[0009] The short-term forecast assessment system includes:
[0010] The data interface module is used to receive forecast data streams from the forecast model and observation data streams from the observation network in real time.
[0011] The dynamic evaluation module is communicatively connected to the data interface module and is used to calculate a set of multidimensional evaluation indicators that reflect the spatiotemporal distribution characteristics of the prediction error based on the forecast data stream and the observation data stream.
[0012] The resource decision module is communicatively connected to the dynamic evaluation module and is used to receive the multi-dimensional evaluation indicators and generate resource scheduling instructions for the next calculation cycle based on preset strategy mapping rules.
[0013] A parallel computing cluster, configured with multiple computing nodes, is used to perform computational tasks for short-term forecast models;
[0014] The optimization scheduling module is communicatively connected to both the resource decision module and the parallel computing cluster. It is used to receive the resource scheduling instructions and dynamically allocate and schedule computing resources in the parallel computing cluster accordingly.
[0015] Furthermore, the multidimensional evaluation index includes at least a spatial structure similarity index and a temporal evolution error index.
[0016] Furthermore, the resource decision-making module has a pre-set strategy mapping rule base, which defines the mapping relationship between multi-dimensional evaluation indicators of different numerical ranges and computing resource quotas of different levels.
[0017] Furthermore, the policy mapping rule base is an adaptive rule base that is updated online through machine learning algorithms.
[0018] A parallel computing optimization method, applicable to any of the short-term forecast evaluation systems described above, wherein the method is executed by the optimization scheduling module, includes the following steps:
[0019] S1. Resource instruction reception: Receive resource scheduling instructions generated based on the current period multi-dimensional evaluation indicators from the resource decision module;
[0020] S2. Computational Task Analysis: Based on the parallelized structural features of the short-term forecast model, it is decomposed into multiple computational sub-tasks that are executed in parallel.
[0021] S3. Dynamic scheduling: According to the resource scheduling instructions, the computing subtasks are dynamically allocated to the corresponding computing nodes in the parallel computing cluster;
[0022] S4. Results Integration and Feedback: Monitor the execution status of each computational subtask, integrate the computational results, and feed back the optimized forecast parameters to the short-term forecast model to initiate the next forecast cycle.
[0023] Furthermore, in the dynamic scheduling of S3, "allocating more computing resources" specifically includes at least one of the following methods:
[0024] S3.1. Add physical computing cores or logical threads to the corresponding computing subtasks;
[0025] S3.2 Switch the numerical calculation mode of the corresponding calculation subtask from the first precision level to the second precision level, wherein the second precision level is higher than the first precision level;
[0026] S3.3 Enable or add graphics processor computing units for the corresponding computing subtasks.
[0027] Furthermore, the dynamic scheduling step also includes a memory optimization step: identifying computational subtasks that are strongly correlated in consecutive forecast periods, and marking and retaining their intermediate computation results in a cache for use by computational tasks with the same or similar initialization conditions in subsequent periods.
[0028] Furthermore, in the result integration and feedback of S4, the "optimized forecast parameters" include: initial field perturbation parameters for ensemble forecasting, selection instructions for physical parameterization schemes, and model integration step size adjustment parameters.
[0029] The beneficial effects of this invention are as follows:
[0030] In the above scheme, through the cooperation of the dynamic evaluation module, resource decision-making module and optimization scheduling module, the real-time forecast evaluation results can be directly transformed into precise resource scheduling instructions, achieving the effect of dynamically allocating computing resources according to forecast uncertainty and realizing the synergistic optimization of computing efficiency and forecast accuracy;
[0031] In the above scheme, by optimizing the parallel computing method executed by the scheduling module and coordinating it with the closed-loop system of the short-term forecast evaluation system, the calculation strategy can be automatically adjusted after each forecast period and fed back to the forecast model. This achieves the effect of enabling the entire system to have self-learning and self-optimization capabilities, and continuously improving the accuracy of short-term forecasts.
[0032] In summary, this device not only effectively solves the problem of rigid resource allocation caused by the disconnect between evaluation and calculation in traditional systems, but also constructs an intelligent feedback loop from evaluation to optimization, providing technical support for the continuous evolution of short-term forecasting capabilities, and has good overall performance. Attached Figure Description
[0033] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention, making other features, objects, and advantages of the invention more apparent. The illustrative embodiments of the invention, along with their descriptions, are used to explain the invention and do not constitute an undue limitation of the invention.
[0034] Furthermore, throughout the accompanying drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the elements are not necessarily drawn to scale.
[0035] In the attached diagram:
[0036] Figure 1 This is a diagram illustrating the architecture and data flow of the short-term forecast assessment system of the present invention.
[0037] Figure 2 This is a flowchart illustrating the parallel computing optimization method of the present invention;
[0038] Figure 3 This is a schematic diagram of the dynamic scheduling process in the parallel computing optimization method of the present invention. Detailed Implementation
[0039] The short-term forecast assessment system and parallel computing optimization method provided by the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more detailed, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.
[0040] It should be noted that the use of terms such as "an embodiment," "an embodiment," "an exemplary embodiment," and "some embodiments" in the specification indicates that the described embodiment may include a specific feature, structure, or characteristic, but not every embodiment necessarily includes that specific feature, structure, or characteristic. Furthermore, when a specific feature, structure, or characteristic is described in connection with an embodiment, implementing such a feature, structure, or characteristic in conjunction with other embodiments (whether explicitly described or not) should be within the knowledge of those skilled in the art.
[0041] Generally, terms can be understood at least partly from their use in context. For example, depending at least partly on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in a singular sense, or a combination of features, structures, or characteristics in a plural sense. Additionally, the term "based on" can be understood not necessarily to convey an exclusive set of factors, but rather, alternatively, depending at least partly on the context, to allow for the presence of other factors that are not necessarily explicitly described.
[0042] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0043] Reference Figure 1-3 As shown, this invention provides a short-term forecast evaluation system and a parallel computation optimization method, wherein:
[0044] The short-term forecast assessment system specifically includes a data interface module, a dynamic assessment module, a resource decision-making module, a parallel computing cluster, and an optimization scheduling module. The data interface module serves as the system's input port, continuously receiving data streams from external sources. These include gridded forecast data streams output from numerical weather prediction models and real-time observation data streams transmitted from observation networks covering the monitoring area (such as weather radar stations, automatic weather stations, traffic flow detectors, or power load monitoring points). In practice, these raw data undergo necessary preprocessing, including format standardization, time alignment, and quality control, before being sent to the subsequent dynamic assessment module.
[0045] In this embodiment, the preprocessing operations described above are mainly as follows: Specifically, because the data formats output by the observation network and the forecast model differ (for example, radar data may be in a specific binary format, while numerical forecast output is in GRIB or NetCDF format, and ground station data is in CSV format), this system will convert and encapsulate all data into a unified and efficient data structure within the system to ensure that the dynamic evaluation module can read and parse it unambiguously. Simultaneously, because there may be slight asynchrony in timestamps between the observation data and the forecast data (e.g., observations are recorded at the hour, while model outputs are at 10-minute intervals), this system will use a high-precision interpolation algorithm (e.g., linear or spline interpolation for continuous fields, and nearest neighbor matching for discrete events) based on the target time of the forecast evaluation to calibrate all data sequences onto the same standard time axis, thereby ensuring the comparability of forecast values and observed values in the time dimension. Finally, the system performs quality control operations, including but not limited to: applying boundary value checks to remove outliers that clearly exceed physically reasonable ranges (such as 300 mm of rainfall per minute); performing spatial consistency analysis to identify and correct or remove isolated erroneous data caused by single-point instrument failures; and normalizing or standardizing the data to eliminate differences in units and numerical ranges between different data sources. Therefore, this preprocessing operation ensures the reliability and consistency of the input data.
[0046] The dynamic evaluation module communicates directly with the data interface module and is used to quantitatively characterize the differences between forecasts and actual data. Based on the received forecast and observation data streams, this module calculates a set of multidimensional evaluation indicators that reflect the spatiotemporal distribution characteristics of prediction errors from different dimensions. In this embodiment, these indicators include at least a spatial structure similarity index (SSIM) to measure the spatial morphological consistency between the forecast and the actual data, and a temporal evolution error index (such as the moving vector error calculated based on optical flow) to quantify the degree of agreement between the prediction system and the actual data in terms of temporal evolution. These indicators together constitute a comprehensive and multi-dimensional evaluation of the performance of the current forecast period.
[0047] The resource decision-making module is connected to the dynamic evaluation module and is used to receive the aforementioned multi-dimensional evaluation indicators. This resource decision-making module has a pre-built strategy mapping rule base, which is essentially a lookup table or decision matrix. It defines the mapping relationship between multi-dimensional evaluation indicators of different numerical ranges and different levels of computing resource quotas. Specifically, when the "temporal evolution error index" of a certain region exceeds threshold A, it is determined that the forecast uncertainty of that region is high, and a "Level 3" computing resource scheduling instruction needs to be triggered. To further improve the system's intelligence level, in this embodiment, the strategy mapping rule base is an adaptive rule base that can be updated online through machine learning algorithms (such as reinforcement learning). Therefore, this system can dynamically adjust the mapping rules based on historical scheduling instructions and the resulting forecast improvement effects, thereby continuously optimizing the decision logic and ensuring that the resource decision-making module ultimately outputs a concrete resource scheduling instruction for the next computing cycle.
[0048] The system's execution layer consists of a parallel computing cluster and an optimization scheduling module. The parallel computing cluster comprises multiple servers or high-performance computing nodes, providing a vast pool of computing resources for executing short-term forecasting model calculations. The optimization scheduling module, acting as the system's brain, communicates with both the resource decision-making module and the parallel computing cluster. It receives resource scheduling instructions and translates them into actual control actions for the parallel computing cluster.
[0049] Specifically, the optimization scheduling module executes a parallel computing optimization method, which includes the following consecutive steps:
[0050] First, in the resource instruction receiving step, the module obtains resource scheduling instructions generated based on the current period assessment results from the resource decision module. Next, in the computation task parsing step, the module analyzes the short-term forecast model expected to be executed next week, scientifically decomposing it into multiple computational subtasks that can be executed independently or collaboratively in parallel, based on its spatial grid partitioning or physical processes.
[0051] Then, the core dynamic scheduling step begins. The optimization scheduling module intelligently allocates these computational subtasks to the corresponding computing nodes in the parallel computing cluster based on resource scheduling instructions. For subtasks in high-uncertainty regions that are explicitly specified in the resource scheduling instructions as requiring priority, the module allocates more computing resources to them. This "allocation of more resources" manifests in several ways at the hardware level. It can be adding physical computing cores or logical threads to the corresponding computational subtask to improve its computational throughput; it can be switching its numerical computation mode from a lower first-precision level (such as single-precision floating-point) to a higher second-precision level (such as double-precision floating-point), sacrificing some computing power for higher numerical precision; or it can be enabling or adding dedicated graphics processing unit (GPU) computing units to the corresponding computational subtask, utilizing its parallel architecture to accelerate specific computations. Furthermore, a memory optimization step can be embedded in this step. The system analyzes data from multiple consecutive forecast periods to identify computational subtasks with strong spatiotemporal correlations (e.g., continuously developing convective cells in weather forecasts), and marks and retains their intermediate computation results in GPU shared memory or node cache. When calculation tasks with the same or similar initialization conditions occur in subsequent forecast periods, these cached results can be directly called, which greatly reduces redundant calculations and data transmission overhead, and improves overall efficiency.
[0052] Finally, in the results integration and feedback step, the optimization scheduling module monitors the execution status of all computational subtasks and integrates the computational results of each node to generate the final optimized forecast product. Simultaneously, the system sends key parameters generated during the optimization process back to the short-term forecast model itself as feedback information. These optimized forecast parameters specifically include, but are not limited to: initial field perturbation parameters used to initialize ensemble forecasts to better characterize model uncertainties; selection instructions for physical parameterization schemes chosen for different regions; or model integration step size adjustment parameters dynamically adjusted to balance accuracy and speed.
[0053] Therefore, through this feedback mechanism, the forecasting model can self-adjust based on the evaluation and optimization results of the current cycle at the beginning of the next cycle, thus forming a continuous improvement closed loop of "evaluation-decision-optimization-forecasting-re-evaluation", ultimately achieving a spiral increase in the accuracy and efficiency of the short-term forecasting system.
[0054] Working principle: In practical use, after the system is started, the data interface module continuously imports forecast and observation data. After preprocessing, the dynamic evaluation module calculates multi-dimensional evaluation indicators to complete the accurate diagnosis of the current forecast effect.
[0055] Subsequently, the resource decision-making module uses its built-in (and adaptively updatable) strategy mapping rules to make intelligent decisions based on the diagnostic results and generate resource scheduling instructions. These instructions mainly include the weak links in the forecasting system and optimization directions.
[0056] Next, the optimization scheduling module acts as the execution hub. By analyzing the predicted tasks and executing parallel computing optimization methods, it transforms decisions into actions: In the dynamic scheduling step, it accurately allocates resources and optimizes memory for the parallel computing cluster according to instructions, ensuring that the superior computing power is concentrated on the key computing tasks that need the most improvement.
[0057] Ultimately, the optimization results are fed back to the forecast model in the form of parameters during the results integration and feedback step, thereby directly optimizing the initial conditions or calculation process for the next round of forecasts.
[0058] This closed-loop process repeats itself, enabling the entire system to dynamically adjust and continuously improve based on its real-time performance. Ultimately, it achieves a virtuous cycle of evaluation-guided optimization, optimization-enhanced forecasts, and forecasts being evaluated again, fundamentally solving the technical problem of the disconnect between evaluation and calculation.
[0059] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
[0060] It should be noted that this invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary confusion regarding the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.
[0061] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A nowcast evaluation system, characterized by, include: The data interface module is used to receive forecast data streams from the forecast model and observation data streams from the observation network in real time. The dynamic evaluation module is communicatively connected to the data interface module and is used to calculate a set of multidimensional evaluation indicators that reflect the spatiotemporal distribution characteristics of the prediction error based on the forecast data stream and the observation data stream. The resource decision module is communicatively connected to the dynamic evaluation module and is used to receive the multi-dimensional evaluation indicators and generate resource scheduling instructions for the next calculation cycle based on preset strategy mapping rules. A parallel computing cluster, configured with multiple computing nodes, is used to perform computational tasks for short-term forecast models; The optimization scheduling module is communicatively connected to both the resource decision module and the parallel computing cluster. It is used to receive the resource scheduling instructions and dynamically allocate and schedule computing resources in the parallel computing cluster accordingly.
2. The short-term forecast assessment system according to claim 1, characterized in that: The multidimensional evaluation indicators include at least spatial structure similarity indicators and temporal evolution error indicators.
3. The short-term forecast assessment system according to claim 1, characterized in that: The resource decision-making module has a pre-set strategy mapping rule base, which defines the mapping relationship between multi-dimensional evaluation indicators in different numerical ranges and computing resource quotas of different levels.
4. The short-term forecast assessment system according to claim 3, characterized in that: The policy mapping rule base is an adaptive rule base that is updated online using machine learning algorithms.
5. A parallel computation optimization method, applicable to the short-term forecast evaluation system according to any one of claims 1-4, characterized in that: The method is executed by the optimized scheduling module and includes the following steps: S1. Resource instruction reception: Receive resource scheduling instructions generated based on the current period multi-dimensional evaluation indicators from the resource decision module; S2. Computational Task Analysis: Based on the parallelized structural features of the short-term forecast model, it is decomposed into multiple computational sub-tasks that are executed in parallel. S3. Dynamic scheduling: According to the resource scheduling instructions, the computing subtasks are dynamically allocated to the corresponding computing nodes in the parallel computing cluster; S4. Results Integration and Feedback: Monitor the execution status of each computational subtask, integrate the computational results, and feed back the optimized forecast parameters to the short-term forecast model to initiate the next forecast cycle.
6. The parallel computing optimization method according to claim 5, characterized in that: In the dynamic scheduling of S3, dynamic allocation specifically includes at least one of the following methods: S3.1 Add physical computing cores or logical threads to the corresponding computing subtasks; S3.2 Switch the numerical calculation mode of the corresponding calculation subtask from the first precision level to the second precision level, wherein the second precision level is higher than the first precision level; S3.3 Enable or add graphics processor computing units for the corresponding computing subtasks.
7. The parallel computing optimization method according to claim 5, characterized in that: The dynamic scheduling step also includes a memory optimization step: identifying computational subtasks that are strongly correlated in consecutive forecast periods, and marking and retaining their intermediate computation results in a cache for use by computational tasks with the same or similar initialization conditions in subsequent periods.
8. The parallel computing optimization method according to claim 5, characterized in that: In the result integration and feedback of S4, the optimized forecast parameters include: initial field perturbation parameters for ensemble forecasting, physical parameterization scheme selection instructions, and model integration step size adjustment parameters.
Citation Information
Patent Citations
Industrial salt production statistical method and system
CN117252447A
Grain storage and transportation disorder plan autonomous scheduling system and method
CN119378962A