Real-time calibration method and system for edge-side meteorological prediction model, electronic equipment and storage medium

Abstract

The invention relates to the technical field of computers, and discloses a real-time calibration method and system for an edge-side weather prediction model, electronic equipment and a storage medium, and the method comprises the steps: obtaining multi-dimensional weather time sequence data, inputting the data into a neural network model, extracting spatial-temporal features, and outputting an initial prediction value; decomposing a weight matrix in the neural network model into an amplitude component and a directional component, performing low-rank matrix updating on the directional component to obtain an updated directional component, calculating a calibration weight according to the amplitude component and the updated directional component, performing deviation correction on the initial predicted value in combination with local observation data, and generating a calibrated predicted value; historical samples and current samples are dynamically managed based on prediction uncertainty indexes, an anti-forgetting training target is constructed in combination with weight regularization constraints, and low-rank matrix parameters are updated to cope with data distribution changes; according to the invention, the reliability of real-time weather prediction service can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computers, and in particular to a real-time calibration method and system for an edge-side weather prediction model, an electronic device and a storage medium. BACKGROUND

[0002] Weather prediction plays a vital role in modern production and life. Traditional weather prediction models are usually complex deep neural networks deployed on cloud computing centers or high-performance server clusters. This centralized processing mode has several inherent defects: end-to-end transmission delay and server-side computation queuing can seriously affect the real-time performance of prediction results, making it difficult to meet the urgent needs of edge-side devices (such as autonomous vehicles and portable weather stations) for immediate decision-making; continuously uploading sensitive local observation data to the cloud poses a risk of privacy leakage; and the operation of the model is highly dependent on stable network connections, making it difficult to work normally in remote areas with poor network conditions or interruptions.

[0003] To solve the above problems, existing technologies attempt to deploy lightweight prediction models directly on edge devices. However, edge devices are limited by their computing power, memory capacity, and energy consumption budget, and cannot undertake the full fine-tuning process of traditional models. In addition, general models will produce prediction biases in specific edge environments (such as specific valleys and urban canyons) due to local climate characteristics, sensor bias, or seasonal changes (referred to as "concept drift"), and this bias will accumulate over time, leading to continuous degradation of model performance.

[0004] Therefore, the present application provides a real-time calibration method for an edge-side weather prediction model to solve the above technical problems. SUMMARY

[0005] The purpose of the present application is to provide a real-time calibration method, system, electronic device and storage medium for an edge-side weather prediction model to solve the technical problem that in the prior art, there is a lack of effective mechanisms for real-time, online, and continuous calibration of edge models under strict resource constraints, making it difficult to balance calibration accuracy, resource efficiency, and long-term stability.

[0006] To solve the above technical problems, the present application provides a real-time calibration method for an edge-side weather prediction model, comprising: Obtaining multi-dimensional weather time series data, inputting it into a compressed and optimized lightweight hybrid neural network model, extracting spatio-temporal features and outputting initial prediction values; decompose the weight matrix in the lightweight hybrid neural network model into a magnitude component and a direction component, update the direction component by a low-rank matrix, obtain an updated direction component, calculate a calibration weight according to the magnitude component and the updated direction component, combine local observation data to correct the deviation of the initial prediction value, and generate a calibrated prediction value; Based on the prediction uncertainty index, the historical samples and the current samples are dynamically managed, and the weight regularization constraint is combined to construct an anti-forgetting training target, and the low-rank matrix parameters are updated to cope with the change of data distribution; The lightweight hybrid neural network model and the low-rank matrix are quantized respectively to reduce the consumption of calculation and storage resources; The prediction error statistics are monitored, and the sample storage size and the parameter update rate are dynamically adjusted based on the error change to maintain the stability of the calibration performance.

[0007] In some embodiments, multi-dimensional meteorological time series data is obtained, input into a compressed and optimized lightweight hybrid neural network model, spatial and temporal features are extracted, and an initial prediction value is output, further comprising: The multi-dimensional meteorological time series data is received, including multiple meteorological element observation values at multiple consecutive time steps; The input multi-dimensional meteorological time series data is convolved by a one-dimensional convolution layer, multiple local spatiotemporal feature patterns at different time scales are extracted using multiple different size convolution kernels, and a feature sequence containing rich feature information is output; The obtained feature sequence is input into a long short-term memory network to capture long-term dependence and temporal dynamic change rules; The initial prediction value is output, including prediction results of multiple meteorological elements at one or more future time steps.

[0008] In some embodiments, the lightweight hybrid neural network model is compressed and optimized, further comprising: The trained lightweight hybrid neural network model is analyzed for weights, and connections with absolute values below a preset threshold are identified and removed; The remaining weights of the pruned model are subjected to low-bit quantization processing, and the original precision floating-point weight values are mapped to an integer representation range; The compressed and optimized lightweight hybrid neural network model is obtained.

[0009] In some embodiments, the weight matrix in the lightweight hybrid neural network model is decomposed into a magnitude component and a direction component, the direction component is updated by a low-rank matrix, an updated direction component is obtained, a calibration weight is calculated according to the magnitude component and the updated direction component, the initial prediction value is corrected for deviation in combination with local observation data, and a calibrated prediction value is generated, further comprising: decomposing each weight matrix in the light-weight hybrid neural network model into two components, an amplitude component representing weight magnitude and a direction component representing weight direction information; updating the direction component by low-rank matrix update, adjusting and optimizing the direction component by introducing the product of two low-rank matrices; calculating calibration weights according to the amplitude component and the updated direction component; calculating a calibrated prediction value by forward propagation based on the calibration weights, the initial prediction value and local observation data.

[0010] In some embodiments, the low-rank matrix update of the direction component further comprises: quantizing the two low-rank matrices used to update the direction component by low-bit, reducing numerical representation accuracy; mapping each element value in the low-rank matrix from the original precision floating-point representation to the integer representation interval.

[0011] In some embodiments, based on the prediction uncertainty index, dynamically manage historical samples and current samples, construct an anti-forgetting training target combining weight regularization constraints, update low-rank matrix parameters to cope with data distribution changes, further comprising: establishing and maintaining a sample buffer for storing historical observation data and corresponding timestamp information; updating the sample content in the sample buffer according to the prediction uncertainty index of the newly arrived sample using an uncertainty priority strategy; real-time calculation of the prediction uncertainty of the newly arrived sample and the uncertainty comparison result of the existing sample in the sample buffer, to decide whether to store the new sample in the sample buffer; combining the newly arrived sample data and the historical sample data saved in the sample buffer, and adding a weight regularization constraint term to jointly construct an anti-forgetting training target function for updating the low-rank matrix parameters.

[0012] In some embodiments, monitor the prediction error statistics, dynamically adjust the sample storage size and parameter update rate based on error changes to maintain stable calibration performance, further comprising: continuously calculating and recording the error statistical characteristics between the calibrated prediction value and the actual observation value; periodically determining whether the error statistical characteristics exceed a preset threshold, the preset threshold being determined based on historical error data distribution characteristics; when the error statistical characteristics exceed the preset threshold, dynamically adjusting the sample storage size to increase the capacity of the sample buffer; simultaneously adjusting the learning rate of parameter update to improve the speed and ability of the model to adapt to data changes.

[0013] Based on the same concept, the application also provides a real-time calibration system for an edge-side weather prediction model, comprising: An initial prediction value output module configured to obtain multi-dimensional weather time series data, input the data into a compressed and optimized lightweight hybrid neural network model, extract spatio-temporal features, and output initial prediction values; A calibrated prediction value output module configured to decompose a weight matrix in the lightweight hybrid neural network model into a magnitude component and a direction component, update the direction component to obtain an updated direction component, calculate calibration weights based on the magnitude component and the updated direction component, correct the initial prediction values based on local observation data, and generate calibrated prediction values; A low-rank matrix parameter update module configured to dynamically manage historical samples and current samples based on a prediction uncertainty index, construct an anti-forgetting training target in combination with weight regularization constraints, and update low-rank matrix parameters to cope with changes in data distribution; A quantization processing module configured to perform quantization processing on the lightweight hybrid neural network model and the low-rank matrix respectively, and reduce the consumption of computing and storage resources; A prediction error statistics module configured to monitor prediction error statistics, dynamically adjust sample storage size and parameter update rate based on error changes, and maintain stable calibration performance.

[0014] Based on the same concept, the application also provides an electronic device, comprising a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of a real-time calibration method for an edge-side weather prediction model.

[0015] Based on the same concept, the application also provides a computer-readable storage medium storing a computer program executable by an electronic device, and when the computer program runs on the electronic device, the electronic device executes the steps of a real-time calibration method for an edge-side weather prediction model.

[0016] Compared with the prior art, the application has the following beneficial effects: The application discloses a real-time calibration method and system for an edge-side weather prediction model, an electronic device, and a storage medium, which improves parameter efficiency and resource adaptability. By adopting a parameter efficient fine-tuning framework and a rank adaptive technology, only a very low-rank parameter matrix needs to be updated to achieve effective calibration. In combination with layered quantization of a core model and a low-rank matrix, the calculation complexity, memory occupation, and energy consumption are reduced, so that a complex real-time calibration task can be stably run on an edge device with highly limited computing and storage resources.

[0017] Real-time calibration is achieved within milliseconds, ensuring the timeliness of prediction. Since the amount of parameters to be updated is extremely small and low-precision quantization is performed, the computational overhead of the calibration process is extremely low, and the time consumption of single-sample update can be controlled within milliseconds. This enables the model to immediately adjust parameters and output predictions based on the latest observation data, effectively meeting the stringent requirements of edge applications for ultra-low latency and overcoming the disadvantages of large transmission delay of traditional cloud models.

[0018] Enhancing the long-term robustness of the model in dynamic environments, the introduction of a continuous learning closed loop that integrates uncertainty management includes a dynamic sample buffer update mechanism based on uncertainty and an anti-forgetting training strategy, enabling the model to intelligently filter and remember high-value historical data, effectively alleviating the problem of catastrophic forgetting. Combined with real-time monitoring and drift detection mechanisms for prediction errors, the system can automatically detect changes in data distribution and dynamically adjust learning strategies to ensure the stability and accuracy of the calibration system during long-term operation.

[0019] Balancing performance and power consumption, the lightweight model structure and low-precision quantization calculation are well suited to the architectural characteristics of edge neural processing units (NPUs), facilitating the use of hardware acceleration units to further improve computational efficiency, achieving inference speed improvement while controlling peak memory usage at a very low level, providing a foundation for efficient deployment and application on various edge chip platforms. BRIEF DESCRIPTION OF DRAWINGS

[0020] Other features, objects, and advantages of the application will become more apparent with reference to the following detailed description of non-limiting embodiments, when read together with the accompanying drawings: Figure 1 is a flowchart of the real-time calibration method of an edge-side weather prediction model according to some embodiments of the present application; Figure 2 is a structural diagram of the real-time calibration system of an edge-side weather prediction model according to some embodiments of the present application; Figure 3 is a structural diagram of an electronic device according to some embodiments of the present application; In the figure, 710 is a processor; 720 is a memory; 730 is an input device; and 740 is an output device. DETAILED DESCRIPTION

[0021] In order to make the purposes, technical solutions, and advantages of the present application clearer, the following will further describe the present application in conjunction with the accompanying drawings. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0022] The terminology used in the embodiments of the present application is for the purpose of describing particular embodiments only and is not intended to be limiting of the present application. As used in the description of the embodiments of the present application and the appended claims, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be understood that the term "plurality" generally includes at least two.

[0023] It should be understood that the term "and / or" used herein is only a description of association relationship of associated objects, which means that there can be three relationships, for example, A and / or B can mean that A exists alone, A and B exist together, and B exists alone. In addition, the character " / " in this paper generally represents that the front and rear associated objects are in an "or" relationship.

[0024] It should be understood that although the terms first, second, third, etc. can be used in the embodiments of the present application to describe, these descriptions should not be limited to these terms. These terms are only used to distinguish the description. For example, without departing from the scope of the embodiments of the present application, the first can also be called the second, and similarly, the second can also be called the first.

[0025] Depending on the context, the word "if" as used herein can be interpreted to mean "when" or "while" or "in response to determining" or "in response to detecting". Similarly, depending on the context, the phrase "if it is determined" or "if (a stated condition or event) is detected" can be interpreted to mean "when it is determined" or "in response to determining" or "when (a stated condition or event) is detected" or "in response to detecting (a stated condition or event)".

[0026] It should also be noted that the terms "comprising", "containing", or any other variant thereof are intended to cover non-exclusive inclusion, so that the product or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such product or device. Without more limitation, the element defined by the sentence "comprising a" does not exclude the existence of other identical elements in the product or device including the element.

[0027] It should be particularly noted that the symbols and / or numbers present in the specification, if not marked in the description, are not the figure marks.

[0028] Reference Figure 1 A real-time calibration method of an edge side weather prediction model, comprising: S101, acquiring multi-dimensional weather time series data, inputting to a compressed and optimized lightweight hybrid neural network model, extracting spatio-temporal features and outputting initial prediction values; S102, decompose the weight matrix in the light-weight hybrid neural network model into a magnitude component and a direction component, update the direction component as a low-rank matrix, obtain an updated direction component, calculate a calibration weight according to the magnitude component and the updated direction component, correct the initial prediction value by combining local observation data, and generate a calibrated prediction value; S103, dynamically manage historical samples and current samples based on a prediction uncertainty index, construct an anti-forgetting training target by combining weight regularization constraints, and update low-rank matrix parameters to cope with data distribution changes; S104, quantize the light-weight hybrid neural network model and the low-rank matrix respectively to reduce the consumption of computing and storage resources; S105, monitor prediction error statistics, and dynamically adjust the sample storage size and the parameter update rate based on error changes to maintain stable calibration performance.

[0029] Specifically, in the embodiments of the present application, time series data containing temperature, humidity, air pressure and other multi-dimensional meteorological indicators are obtained and input into a light-weight hybrid neural network model optimized by compression; the model uses a one-dimensional convolutional neural network to extract the local space-time features of the input data, and uses a long short-term memory network to model the time series of the features, and outputs an initial meteorological prediction value. To achieve efficient parameter updating, the weight matrix in the above neural network is decomposed into a scalar component representing the overall magnitude and a unit matrix component representing the direction information; only the direction component is updated as a low-rank matrix, and the original direction component is adjusted by introducing the product of two low-rank matrices, and then the calibrated weight is calculated by combining the magnitude component; the initial prediction value is calculated forward using the weight, and the prediction bias is calculated by a loss function combining real-time local observation data, to generate a calibrated prediction result. To cope with the problem of data distribution changing over time, a sample management strategy based on prediction uncertainty is used to dynamically maintain a historical sample buffer, and samples with high uncertainty are preferentially retained; when updating the low-rank parameters, the new samples and the historical samples in the buffer are combined, and a weight change regularization constraint is introduced to jointly construct an anti-forgetting training target function, to maintain the model's memory of past knowledge. To further reduce resource consumption, the weight parameters in the light-weight hybrid neural network model and the low-rank update matrix are quantized respectively, converting the original high-precision floating-point numbers to low-bit-width integer representations, significantly reducing memory occupancy and computational overhead. The statistical distribution of prediction errors is continuously monitored, and when the mean or variance of the errors exceeds a set threshold, the capacity of the sample buffer and the learning rate in the parameter update process are automatically adjusted, thereby realizing rapid adaptation to environmental changes and maintaining the long-term stability of the calibration system.

[0030] For example, the method is deployed in a certain edge weather station, which collects temperature, humidity and pressure data every hour, forming a 3-dimensional time series input to a pruned and 8-bit quantized lightweight CNN-LSTM model. The model contains 10 one-dimensional convolution kernels and LSTM units with a hidden layer size of 20. The output is the initial prediction of temperature and humidity 1 hour in the future. Assuming the initial prediction is temperature 25°C and humidity 70%. Decompose a weight matrix in the model to get an amplitude component of 1 unit matrix. Perform a rank-4 low-rank update on the unit matrix, that is, introduce two low-rank matrices matched in dimension to multiply and superimpose on the original unit matrix, and then multiply it with the amplitude component to get the new weight. Using the new weight and the actual measured temperature of 24°C and humidity of 72% from the local sensor, calculate the bias loss to obtain the calibrated prediction value, for example, temperature 24.5°C and humidity 71.5%. The sample buffer capacity is set to 50. The prediction uncertainty of the new sample is obtained by calculating the prediction variance. If the current sample uncertainty is high, it will be added to the buffer. When updating, 10 groups of historical samples are randomly selected from the buffer to calculate the loss together with the current sample, and a weight two-norm regularization term is added to suppress parameter mutation. Quantize all weights in the model to 8 bits and quantize low-rank matrices to 4 bits. Quantize the floating-point weight to an integer between 0 and 255 to reduce storage overhead. Calculate the mean and variance of the prediction error every 24 hours. If the mean error is found to have risen continuously by more than 5% of the threshold, temporarily increase the buffer capacity to 100, and increase the learning rate by 50% to accelerate the model to adapt to the new data pattern, thereby maintaining the prediction accuracy.

[0031] In some applications, multi-dimensional meteorological time series data is obtained, input into a compressed and optimized lightweight hybrid neural network model, and spatial and temporal features are extracted and initial prediction values are output. The method includes receiving the multi-dimensional meteorological time series data, including multiple meteorological element observation values at multiple consecutive time steps; performing convolution operation on the input multi-dimensional meteorological time series data through a one-dimensional convolution layer, using multiple convolution kernels of different sizes to extract local spatio-temporal feature patterns at multiple time scales, and outputting a feature sequence containing rich feature information; inputting the obtained feature sequence into a long short-term memory network to capture long-term dependence and time series dynamic change law; and outputting the initial prediction value, including multiple meteorological element prediction results at one or more future time steps.

[0032] It can be understood that multi-dimensional meteorological time series data is received from a sensor, which contains multiple meteorological element observations collected at multiple consecutive time steps, forming a complete time series input. The time series data is input into a compressed and optimized lightweight hybrid neural network model, which integrates a one-dimensional convolution layer and a long short-term memory network. The one-dimensional convolution layer is used to perform convolution operation on the input time series data, in which multiple convolution kernels with different time widths are used to extract local spatio-temporal feature patterns at different time scales, thereby outputting a feature sequence containing multi-scale feature information. The obtained feature sequence is input into the long short-term memory network, which uses its gating mechanism to capture the long-term dependence relationship and time series dynamic change law existing in the meteorological data. The long short-term memory network outputs an initial prediction value, which contains the prediction results of multiple meteorological elements at one or more future time steps, providing a basis for subsequent calibration operations.

[0033] For example, a certain edge computing node receives temperature, humidity and pressure data collected every hour within the past 24 hours, forming an input matrix with a time step number of 24 and a feature dimension of 3. The matrix is input into a lightweight hybrid neural network that has been compressed by pruning and 8-bit quantization. The one-dimensional convolution layer of the model uses three convolution kernels with sizes of 2, 3 and 4 to extract short-term to medium-term local features, and outputs a feature sequence containing 22 time steps. The feature sequence is input into an LSTM network with 16 hidden units. The LSTM captures long-term dependencies of more than 12 hours through its internal gating state, and outputs temperature and humidity prediction values for the next 3 hours, for example, the first hour prediction value is temperature 25.2°C and humidity 68%, the second hour prediction value is temperature 24.8°C and humidity 71%, and the third hour prediction value is temperature 24.5°C and humidity 73%.

[0034] In some applications, the lightweight hybrid neural network model is compressed and optimized, including weight analysis of the trained lightweight hybrid neural network model, identifying and removing connections with weights less than a preset threshold; low-bit quantization processing of the remaining weights of the pruned model, mapping the original precision floating-point weight values to the integer representation range; obtaining the compressed and optimized lightweight hybrid neural network model.

[0035] It can be understood that the weight analysis is performed on the trained lightweight hybrid neural network model, all weight connections in the network are scanned and the absolute values thereof are calculated, and all connections with an absolute value lower than a preset threshold are identified. The model pruning operation is performed to remove the identified low-importance connections from the network, thereby reducing the parameter quantity and the computational complexity of the model. Then, low-bit quantization processing is performed on the remaining weights in the pruned model, the original high-precision floating-point weight values are converted to the low-bit integer representation range through a linear mapping function, and the storage space requirement of a single weight value is reduced. The compressed and optimized lightweight hybrid neural network model is obtained, which reduces the computational resource consumption and the storage space requirement while maintaining the original prediction performance, and can be deployed on a resource-constrained edge computing device.

[0036] For example, the lightweight hybrid neural network model comprising convolutional layers and LSTM layers is compressed and optimized, all weights in the model are analyzed, an absolute threshold of 0.05 is set, and all weight connections with an absolute value lower than the threshold are removed, so that the model parameter quantity is reduced from the original 100 KB to 60 KB. The remaining weights after pruning are subjected to 8-bit quantization processing, the original 32-bit floating-point weight values are mapped to the integer range of 0 to 255 through linear transformation, for example, the floating-point weight value -0.34 is quantized to the integer 45, and 0.78 is quantized to 200. After compression and optimization, the obtained model volume is reduced to 25 KB, the peak memory occupancy is reduced to less than 15 MB, and the deployment efficiency on the edge device is improved.

[0037] In some applications, the weight matrix in the lightweight hybrid neural network model is decomposed into an amplitude component and a direction component, the direction component is subjected to low-rank matrix updating to obtain an updated direction component, the calibration weight is calculated according to the amplitude component and the updated direction component, the deviation correction is performed on the initial prediction value in combination with the local observation data to generate a calibrated prediction value, including decomposing each weight matrix in the lightweight hybrid neural network model into two components, an amplitude component representing the weight amplitude and a direction component representing the weight direction information; the direction component obtained by decomposition is subjected to low-rank matrix updating, and the direction component is adjusted and optimized by introducing the product of two low-rank matrices; the calibration weight is calculated according to the amplitude component and the updated direction component; and the calibrated prediction value is calculated by forward propagation based on the calibration weight, the initial prediction value and the local observation data.

[0038] It can be understood that each weight matrix in the lightweight hybrid neural network model is decomposed into two components: an amplitude component representing the overall amplitude of the weight and a unitized direction component representing the direction information of the weight. The direction component obtained by decomposition is updated by low-rank matrix, two specific rank matrices are introduced and their product is calculated, the product is added to the original direction component to adjust and optimize the direction component, and the updated direction component is obtained. The calibration weight is calculated by matrix multiplication according to the amplitude component unchanged and the updated direction component. Based on the newly calculated calibration weight, the initial prediction value of the model output and the real-time acquired local observation data, a more accurate calibrated prediction value is calculated by forward propagation, and the deviation correction of the initial prediction result is completed.

[0039] For example, a weight matrix with a dimension of 100x50 in the model is decomposed, and the amplitude component is calculated to be 2.5 and the direction component is a 100x50 unit matrix. The direction component is updated by low-rank with a rank of 4, two low-rank matrices with dimensions of 100x4 and 4x50 are introduced, their product is calculated and added to the original direction component. The calibration weight matrix is calculated according to the amplitude component 2.5 and the updated direction component. The new weight is used to forward calculate the initial prediction value of temperature 25.2℃ and humidity 68%, and combined with the measured temperature 24.5℃ and humidity 72% of the local sensor, the calibrated prediction value of temperature 24.8℃ and humidity 70.5% is generated by minimizing the mean square error loss.

[0040] In some applications, the low-rank matrix update of the direction component also includes low-bit quantization processing of the two low-rank matrices used to update the direction component to reduce the numerical representation precision; each element value in the low-rank matrix is mapped from the original precision floating point representation to the integer representation interval.

[0041] It can be understood that in the low-rank matrix update process, the two low-rank matrices used to update the direction component are further subjected to low-bit quantization processing, which reduces the consumption of computing resources by reducing the numerical representation precision of the matrix elements. Each element value in the low-rank matrix is converted from the original high-precision floating point representation to the low-bit integer representation interval by a linear mapping function, reducing the storage space requirement of the matrix and the memory occupation in the calculation process, while maintaining the numerical distribution characteristics of the matrix, ensuring the effectiveness and stability of the low-rank update operation, and efficiently completing the optimization and adjustment of the direction component in a resource-limited environment.

[0042] For example, two low-rank matrices used to update the direction component are quantized by 4 bits, and each element value in the first low-rank matrix with a dimension of 100x4 and the second low-rank matrix with a dimension of 4x50 is mapped from the original 32-bit floating-point representation to the integer interval of 0 to 15. Specifically, the floating-point value -0.32 is converted to the integer 3 by the quantization formula, 0.75 is converted to the integer 12, and 1.28 is converted to the integer 15. After 4-bit quantization, the storage space of each matrix element is reduced from 32 bits to 4 bits, and the total storage requirement of the two low-rank matrices is reduced from 18,000 bits to 2,250 bits, reducing the storage overhead by 87.5% and improving the calculation efficiency on the edge device.

[0043] In some applications, the historical samples and the current samples are dynamically managed based on the prediction uncertainty indicators, an anti-forgetting training target is constructed by combining weight regularization constraints, and the low-rank matrix parameters are updated to cope with data distribution changes, including establishing and maintaining a sample buffer for storing historical observation data and corresponding timestamp information; according to the prediction uncertainty indicators of the newly arrived samples, the sample content in the sample buffer is updated using an uncertainty priority strategy; the prediction uncertainty of the newly arrived samples is calculated in real time and compared with the uncertainty of the existing samples in the sample buffer to decide whether to store the new samples in the sample buffer; the newly arrived sample data and the historical sample data saved in the sample buffer are combined, and a weight regularization constraint term is added to jointly construct an anti-forgetting training target function for updating the low-rank matrix parameters.

[0044] It can be understood that the sample buffer is established and maintained for storing historical observation data and corresponding timestamp information to form a valuable historical data set. According to the prediction uncertainty indicator values of the newly arrived samples, the sample content in the sample buffer is updated using an uncertainty priority strategy, and the sample data with higher prediction uncertainty is preferentially retained. The prediction uncertainty value of the newly arrived samples is calculated in real time and compared with the uncertainty indicators of the existing samples in the sample buffer for analysis, and based on the comparison result, it is decided whether to store the new samples in the sample buffer. The newly arrived sample data and the historical sample data saved in the sample buffer are combined, and a weight regularization constraint term is added to jointly construct an anti-forgetting training target function, and the low-rank matrix parameters are updated by optimizing the target function, so that the model can effectively cope with data distribution changes while avoiding forgetting existing knowledge.

[0045] For example, a sample buffer with a capacity of 100 is established to store temperature and humidity observation data collected in the past 200 hours and corresponding timestamps. The prediction uncertainty of a new sample is obtained by calculating the prediction variance, and the uncertainty of the current new sample is 0.85, and the lowest uncertainty value in the buffer is 0.35. Since the uncertainty of the new sample is higher, the uncertainty priority strategy is used to add it to the buffer, and the sample with an uncertainty of 0.35 is removed. When the model is updated, 20 historical samples are randomly selected from the buffer to calculate the mean square error loss together with the current new sample, and a weight change two-norm regularization term is added, with a regularization coefficient of 0.01. The rank-4 low-rank matrix parameters are updated by the gradient descent algorithm, and the learning rate is set to 0.001.

[0046] In some applications, the prediction error statistics are monitored, and the sample storage size and parameter update rate are dynamically adjusted based on error changes to maintain stable calibration performance. This includes continuously calculating and recording error statistical properties between calibrated prediction values and actual observation values; periodically determining whether the error statistical properties exceed a preset threshold, which is determined based on historical error data distribution characteristics; when the error statistical properties exceed the preset threshold, dynamically adjusting the sample storage size to increase the capacity of the sample buffer; and adjusting the learning rate of the parameter update to improve the speed and ability of the model to adapt to data changes.

[0047] It can be understood that the error statistical properties between the calibrated prediction values and the actual observation values are continuously calculated and recorded, including the mean, variance and trend characteristics of the error over time. Periodically determine whether the error statistical properties exceed the preset threshold determined based on the historical error data distribution characteristics, which reflects the error fluctuation range. When the error statistical properties exceed the preset threshold, it indicates that there may be changes in data distribution or model performance degradation, at which time the sample storage size is dynamically adjusted to increase the capacity of the sample buffer to accommodate more diverse data samples. At the same time, the learning rate in the parameter update process is adjusted accordingly to improve the model's speed and learning ability to adapt to data changes, ensuring long-term stable prediction performance when facing environmental changes.

[0048] For example, the error between the predicted temperature and the actual observed temperature is continuously calculated, and the average error of the past 24 hours is 0.5°C with a variance of 0.2. When the current average error is monitored to rise to 1.2°C and exceed the preset threshold of 0.8°C, it is determined that an anomaly has occurred. The sample buffer capacity is dynamically adjusted from the original 100 samples to 150 samples to retain more historical data features. At the same time, the parameter update learning rate is increased from 0.001 to 0.0015 to enhance the model's adaptability. After 12 hours of operation, the average error falls to 0.6°C, returning to a stable state, at which point the buffer capacity is adjusted back to 120 samples and the learning rate is restored to 0.0012, maintaining optimal performance.

[0049] Another embodiment of a real-time calibration method for an edge-side weather prediction model of the present application is described below: In this embodiment, the model structure includes: The input layer receives multi-dimensional weather time series data (such as temperature, humidity, air pressure, etc., denoted as sequence ; Input layer parameters: : represents the input multi-dimensional weather time series data sequence. Among them, is the number of time steps (such as the number of observed hours, days), is the feature vector of the th time step.

[0050] : The specific observation data of each time step, is the feature dimension (such as the number of weather indicators such as temperature, humidity, air pressure, The larger the value, the more indicators monitored, and the more complex the data in the "feature" dimension.

[0051] CNN feature extraction layer: extract local spatio-temporal features through 1 convolution kernel ( convolution window size) and output , where: ( activation function, bias); CNN feature extraction layer parameters: : Number of convolution kernels (i.e., number of extracted feature channels). The larger the value, the more diverse local patterns can be captured, but the computational load will also increase.

[0052] : Convolution window size (length of sliding window in time dimension). For example =3 means each convolution operation covers 3 consecutive time steps of meteorological data, aiming to capture short-term dependencies (e.g., temperature fluctuation features within 3 hours).

[0053] : Convolution kernel weight matrix. Dimensions are (number of convolution kernels) x (window size) x (input feature dimension), each kernel is responsible for extracting a local feature pattern (e.g., "high temperature - humidity drop" association pattern).

[0054] : Bias term for convolution operation, used to adjust the baseline of convolution output, increasing model flexibility.

[0055] : Activation function (e.g., ReLU, sigmoid), used to introduce non-linear transformation, enabling the model to capture complex feature relationships (e.g., non-linear association between temperature and atmospheric pressure).

[0056] : Output feature sequence of CNN layer. Due to convolution window sliding, the number of output time steps is (shorter than the input sequence by time steps).

[0057] : Feature vector at the th time step in the feature sequence, calculated by convolution kernel and local window of input data (formula ), containing local key features within the time window.

[0058] LSTM time series prediction layer: input feature sequence into LSTM unit, capture long-term dependencies through gating mechanism, output initial prediction value ( is the prediction dimension). LSTM state update formula is: ; ( are input gate, forget gate, output gate respectively, is cell state, is hidden state, For the weight matrix, For the bias).

[0059] Temporal prediction layer parameters: : The output of the LSTM layer, i.e., the initial prediction value. is the prediction dimension (e.g., if the temperature and humidity in the next 1 hour need to be predicted, then = 2).

[0060] LSTM state update formula parameters: The LSTM dynamically adjusts the information flow through the input gate, the forget gate, and the output gate. The core parameters are as follows: : are the input gate, the forget gate, and the output gate (all are vectors) of the first time step, respectively. The value range is between (0, 1) (through the sigmoid activation), which controls the "inflow", "retention", and "outflow" of information.

[0061] Input gate : decides which information in the current input feature and the last hidden state needs to be stored in the cell state.

[0062] Forget gate : decides which information in the last cell state needs to be retained (e.g., whether to "forget" the abnormally high temperature a few days ago).

[0063] Output gate : decides which information in the current cell state needs to be output to the hidden state (affecting the final prediction).

[0064] : : the cell state of the first time step ("memory unit"), which stores long-term information and is obtained by superimposing the historical state filtered by the forget gate and the new information filtered by the input gate. ).

[0065] : the candidate cell state, which is calculated from the current input and the last hidden state and contains new candidate information.

[0066] : the first Hidden state of time step, cell state controlled by output gate after tanh activation , is the immediate output of LSTM, used for subsequent prediction or passed to next time step.

[0067] : : weight matrices of input gate, forget gate, output gate, candidate cell state, respectively, used to map input features and previous hidden state to gating signals or candidate states.

[0068] : corresponding bias terms, used to adjust the baseline of gating signals.

[0069] : Element-wise multiplication (Hadamard product), used for element-wise filtering or stacking information (e.g., multiplication of forget gate and historical cell state, realizing "selective forgetting").

[0070] Optimization strategy: model pruning (removing connections with weights whose absolute values are less than a threshold ) and 8-bit quantization (mapping weights to interval: , to compress the model volume to the range that edge devices can bear (usually <100MB, depending on the specific case).

[0071] Second layer: DoRA real-time calibration module Decoupled Rank Adaptation (DoRA) technology is used as the core of efficient fine-tuning of parameters. Under the premise of freezing the core model weights, local bias correction is achieved through low-rank updating, balancing calibration accuracy and resource efficiency. Weight decomposition principle: Any weight matrix in the core model is decomposed into amplitude and direction components: , where (F-norm normalized amplitude), (unit direction matrix).

[0072] Note: F-norm normalized amplitude is to first calculate the F-norm (square root of the sum of squares of each element) of the matrix (or vector), then divide each element of the original matrix (or vector) by the F-norm, thus scaling its amplitude to the unit F-norm standardization operation.

[0073] DoRA only updates the direction component with low rank, avoiding the weight norm drift problem in LoRA: is a low-rank matrix, is the rank, usually takes 4-8); The final calibrated weights are .

[0074] Calibration process: Input the core model output and the local observation value , learn the bias correction function through the DoRA module , output the calibrated prediction value: ; The formula is explained as follows: Symbol meaning: : The final calibrated prediction value (such as calibrated temperature, humidity, etc. meteorological indicators).

[0075] : The initial prediction value output by the core model (CNN+LSTM hybrid architecture).

[0076] : The bias correction function learned by the DoRA module, used to correct the prediction bias of the core model.

[0077] : The trainable parameters of the DoRA module (low-rank matrix and ), dynamically adjusted through real-time learning of local observation data.

[0078] Formula logic: The calibrated prediction result = the initial prediction value of the core model + the bias correction amount calculated by the DoRA module, where the bias correction function takes the initial prediction of the core model as input, and based on the low-rank parameters of DoRA, calculates the bias value that needs to be corrected, finally realizes the dynamic calibration of the initial prediction.

[0079] Core role: Under the premise of freezing the core model weights, through a small amount of parameters (only <1% of the core model), the local bias correction is realized, taking into account the resource constraints and prediction accuracy requirements of edge devices, supporting real-time updating (single sample update time <10ms).

[0080] Updating method: Real-time updating by minimizing the loss function .

[0081] ​Quantization adaptation: combined with 4-bit quantization (QLoRA variant), quantize low-rank matrices to [0, 15] interval, further reduce memory footprint by zero-point quantization formula ( = zero point, = scale factor) to ensure real-time training on edge devices (single-sample update time < 10ms).

[0082] The explanation of this process is as follows: Core objective: On the basis of freezing core model weights and only real-time calibration through low-rank matrices and (trainable parameters of DoRA module), further compress the memory footprint of and to ensure that edge devices can complete training updates in a very short time (single-sample update time < 10ms), balancing calibration accuracy and resource efficiency.

[0083] 4-bit quantization (QLoRA variant) quantization meaning: compress the numerical values in and that would originally be stored as 32-bit or 16-bit floating-point numbers into 4-bit integers (i.e. possible values), thereby reducing memory footprint to (32-bit -> 4-bit) or (16-bit -> 4-bit).

[0084] Quantization range: map the element values of and to the [0, 15] interval (4-bit integer value range), replacing the original floating-point representation.

[0085] Zero-point quantization: : original floating-point number (some element in low-rank matrix or ).

[0086] : zero point (offset), a constant to be determined, used to adjust the baseline of the quantization range (e.g., map negative numbers to the non-negative interval [0, 15]).

[0087] : scale factor (scale), used to control the mapping ratio of the original value to the quantization interval, the calculation formula is usually (uniformly map the dynamic range of the original value to 16 discrete values).

[0088] Rounding ensures the result is an integer.

[0089] Calculation example: If the original value The range is ,but (Midnight) (Scaling factor). For The quantized value is (falling) Within the interval).

[0090] Functions and advantages: Reduced memory footprint: Low rank moment achieved through 4-bit quantization. and This significantly reduces storage requirements and solves the memory-constrained problem of edge devices.

[0091] Ensuring real-time performance: The smaller number of parameters reduces computational complexity, enabling the DoRA module to update a single sample in less than 10ms, meeting the requirements for real-time calibration at the edge.

[0092] Compatible with QLoRA concept: As a variant of QLoRA (Quantization LoRA), it preserves the distribution characteristics of the original data through zero point and scaling factor during the quantization process, reducing the impact of quantization error on calibration accuracy.

[0093] Third layer: Continuous learning loop A continuous learning mechanism is introduced to address the performance degradation of models caused by "conceptual drift" in meteorological data (such as seasonal variations and sensor aging). Through dynamic sample management and forgetting suppression strategies, long-term stable calibration is achieved.

[0094] Rehearsal buffer design: Maintenance capacity is sample buffer The update adopts an "uncertainty-first" strategy: when a new sample... When it arrives, if the core model predicts uncertainty ( If the threshold is used, then the sample with the lowest uncertainty in the buffer is replaced to ensure that the buffer retains high-value historical data.

[0095] Anti-forgetting training: Each time the DoRA parameters are updated, the loss of the new samples and the loss of the buffer samples are minimized simultaneously, and a weighted regularization term is added to suppress drastic changes in the parameters: ; ( For the loss of new samples, For buffer sample loss, For hyperparameters, is the initial error mean Drift detection: compute the mean and variance of prediction errors by sliding window of prediction errors and variance When ( is the initial error mean ), trigger buffer expansion (temporarily increase to ) and learning rate boost (

[0096] Leverage edge device integrated neural processing units (NPUs) to improve efficiency by: Low-precision computation: map CNN convolutions and LSTM gating operations to NPU's 8-bit arithmetic units, inference speed up by 3-5x Memory reuse: control peak memory footprint to < 32MB by sharing weights across model layers (e.g. LSTM's gating weight shared matrices)

[0097] Software toolchain: Model training: train core model on server side with PyTorch, initialize DoRA module with peft library Device deployment: convert model to TFLite format with TensorFlow Lite for Microcontrollers, implement inference and training logic in C++ Monitoring and updating: edge device periodically upload calibration error statistics to cloud, cloud aggregates and updates DoRA initialization parameters via federated learning, push to device to optimize initial calibration performance

[0098] For the method steps disclosed in the above embodiments, the method steps are described as a combination of a series of actions for the purpose of simple description, but those skilled in the art should know that the embodiments of the present application are not limited by the order of the described actions, because according to the embodiments of the present application, certain steps can be performed in other order or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily necessary for the embodiments of the present application.

[0099] As shown in Figure 2 , the present application also provides a real-time calibration system for an edge side weather prediction model, comprising: An initial prediction value output module 201 configured to obtain multi-dimensional weather time series data, input the data into a compressed and optimized lightweight hybrid neural network model, extract spatio-temporal features and output initial prediction values The post-calibration prediction value output module 202 is configured to decompose a weight matrix in the light-weight hybrid neural network model into an amplitude component and a direction component, update the direction component by a low-rank matrix, obtain an updated direction component, calculate a calibration weight according to the amplitude component and the updated direction component, correct a bias of the initial prediction value by combining local observation data, and generate a post-calibration prediction value. The low-rank matrix parameter updating module 203 is configured to dynamically manage historical samples and current samples based on a prediction uncertainty index, construct an anti-forgetting training target by combining a weight regularization constraint, and update a low-rank matrix parameter to cope with data distribution changes. The quantization processing module 204 is configured to quantize the light-weight hybrid neural network model and the low-rank matrix respectively, and reduce calculation and storage resource consumption. The prediction error statistics module 205 is configured to monitor prediction error statistics, dynamically adjust sample storage size and parameter update rate based on error changes, and maintain stable calibration performance.

[0100] It is worth noting that, although only some basic functional modules are disclosed in the embodiments of the present application, it does not mean that the composition of the system is limited to only the above basic functional modules. On the contrary, the meaning expressed in the embodiments is that on the basis of the above basic functional modules, those skilled in the art can add one or more functional modules to form infinite embodiments or technical solutions in combination with existing technologies. That is, the system is open rather than closed, and the protection scope of the present application claims cannot be limited to the disclosed basic functional modules. At the same time, for the convenience of description, the above device is described as various units and modules. Of course, the functions of the units and modules can be realized in the same software and / or hardware when implementing the present application.

[0101] As shown in Figure 3 The present application also provides an electronic device, comprising: a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory complete mutual communication through the communication bus; the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the real-time calibration method of the edge-side weather prediction model.

[0102] Figure 3 is a structural schematic diagram of an electronic device provided by an embodiment of the present application. As shown in the structure Figure 3 The electronic device provided in the embodiment of the present application includes one or more processors 710 and a memory 720; the processor 710 in the electronic device can be one or more, Figure 3The processor 710 is taken as an example; the memory 720 is configured to store one or more programs; the one or more programs are executed by the one or more processors 710, so that the one or more processors 710 implement the real-time calibration method of the edge-side weather prediction model according to any one of the embodiments of the present application.

[0103] The electronic device can further include an input device 730 and an output device 740.

[0104] The processor 710, the memory 720, the input device 730 and the output device 740 in the electronic device can be connected through a bus or other means, Figure 3 The connection through the bus is taken as an example.

[0105] The memory 720 in the electronic device can be used as a computer readable storage medium to store one or more programs, and the program can be a software program, a computer executable program and a module, such as the program instruction / module corresponding to the real-time calibration method of the edge-side weather prediction model provided in the embodiments of the present application. The processor 710 executes the software program, instruction and module stored in the memory 720, so as to execute various function applications and data processing of the electronic device, that is, to implement the real-time calibration method of the edge-side weather prediction model in the above method embodiments.

[0106] The memory 720 can include a program storage area and a data storage area, wherein the program storage area can store an operating system and at least one application required by a function; the data storage area can store data created according to the use of the electronic device, etc. In addition, the memory 720 can include a high-speed random access memory, and can also include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other non-volatile solid-state memory device. In some examples, the memory 720 can further include a memory remotely arranged relative to the processor 710, which can be connected to the device through a network. Examples of the above network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network and a combination thereof.

[0107] The input device 730 can be used to receive input digital or character information, and generate key signal input related to the user settings and function control of the electronic device. The output device 740 can include a display device such as a display screen.

[0108] The present application also provides a computer readable storage medium storing a computer program executable by an electronic device, when the computer program is run on the electronic device, so that the electronic device executes the steps of the real-time calibration method of the edge-side weather prediction model.

[0109] In particular, a computer storage medium of embodiments of the present application can employ any combination of one or more computer readable medium or media. The computer readable medium can be a computer readable signal medium or a computer readable storage medium. A computer readable storage medium can be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer readable storage medium include the following: an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In this embodiment, the computer readable storage medium can be any tangible medium that contains or stores a program used by an instruction execution system, apparatus, or device to function or to be combined with the instruction execution system, apparatus, or device.

[0110] Finally, it should be noted that the above embodiments are merely used to illustrate the technical solutions of the present application, rather than limit the present application; even though the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions recorded in the above embodiments can be modified, or some or all of the technical features can be replaced equivalently; and these modifications or replacements do not cause the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for real-time calibration of edge-side weather prediction models, characterized in that, The method comprises the following steps: acquiring multi-dimensional meteorological time series data, inputting the data into a compressed and optimized lightweight hybrid neural network model, extracting space-time features and outputting initial prediction values; decomposing the weight matrix in the lightweight hybrid neural network model into amplitude components and direction components, updating the direction components by low-rank matrix, obtaining updated direction components, calculating calibration weights according to the amplitude components and the updated direction components, correcting the initial prediction values by combining local observation data, and generating calibrated prediction values; dynamically managing historical samples and current samples based on prediction uncertainty indicators, constructing an anti-forgetting training target by combining weight regularization constraints, and updating low-rank matrix parameters to cope with changes in data distribution; quantizing the lightweight hybrid neural network model and the low-rank matrix respectively to reduce the consumption of calculation and storage resources; monitoring prediction error statistics, dynamically adjusting the sample storage size and parameter update rate based on error changes to maintain stable calibration performance. 2.The method of calibrating a real-time edge-side weather prediction model according to claim 1, wherein, acquiring multi-dimensional meteorological time series data, inputting the data into a compressed and optimized lightweight hybrid neural network model, extracting space-time features and outputting initial prediction values, further comprising: receiving the multi-dimensional meteorological time series data, including multiple meteorological element observation values at multiple consecutive time steps; performing convolution operation on the input multi-dimensional meteorological time series data through a one-dimensional convolution layer, using multiple convolution kernels of different sizes to extract local space-time feature patterns at multiple time scales, and outputting feature sequences containing rich feature information; inputting the obtained feature sequences into a long short-term memory network to capture long-term dependence and time series dynamic change rules; outputting the initial prediction values, including prediction results of multiple meteorological elements at one or more future time steps. 3.The method of claim 1, wherein, compressing and optimizing the lightweight hybrid neural network model, further comprising: performing weight analysis on the trained lightweight hybrid neural network model, identifying and removing connections with weight absolute values below a preset threshold; performing low-bit quantization processing on the remaining weights of the pruned model, mapping the original precision floating-point weight values to the integer representation range; obtaining the compressed and optimized lightweight hybrid neural network model.

4. The method of claim 1, wherein, decomposing the weight matrix in the lightweight hybrid neural network model into amplitude components and direction components, updating the direction components by low-rank matrix, obtaining updated direction components, calculating calibration weights according to the amplitude components and the updated direction components, correcting the initial prediction values by combining local observation data, and generating calibrated prediction values, further comprising: decomposing each weight matrix in the lightweight hybrid neural network model into two components, an amplitude component representing the weight amplitude and a direction component representing the weight direction information; updating the direction components obtained by decomposition by low-rank matrix, adjusting and optimizing the direction components by introducing the product of two low-rank matrices; calculating calibration weights according to the amplitude components and the updated direction components; calculating calibrated prediction values by forward propagation based on the calibration weights, the initial prediction values and local observation data.

5. The method of real-time calibration of a nowcast model for edge weather according to claim 4, characterized in that, updating the direction components by low-rank matrix further comprises: The two low-rank matrices for updating the direction component are subjected to low-bit quantization processing, reducing the numerical representation precision; Each element value in the low-rank matrix is mapped from the original precision floating-point representation to an integer representation interval.

6. The method of real-time calibration of a nowcast model for edge weather according to claim 1, wherein, Based on the prediction uncertainty index, the historical samples and the current samples are dynamically managed, and the anti-forgetting training target is constructed by combining the weight regularization constraint to update the low-rank matrix parameters to cope with the data distribution changes, further comprising: A sample buffer is established and maintained for storing historical observation data and corresponding timestamp information; According to the prediction uncertainty index of the newly arrived sample, the sample content in the sample buffer is updated using the uncertainty priority strategy; The prediction uncertainty of the newly arrived sample is calculated in real time and compared with the uncertainty of the existing samples in the sample buffer to decide whether to store the new sample in the sample buffer; The newly arrived sample data and the historical sample data saved in the sample buffer are combined, and a weight regularization constraint term is added to jointly construct an anti-forgetting training target function for updating the low-rank matrix parameters.

7. The method of real-time calibration of a nowcast model for edge weather according to claim 6, characterized in that, The prediction error statistics are monitored, and the sample storage size and the parameter update rate are dynamically adjusted based on the error changes to maintain stable calibration performance, further comprising: The error statistical characteristics between the calibrated prediction values and the actual observation values are continuously calculated and recorded; It is regularly judged whether the error statistical characteristics exceed a preset threshold, and the preset threshold is determined based on the historical error data distribution characteristics; When the error statistical characteristics exceed the preset threshold, the sample storage size is dynamically adjusted to increase the capacity of the sample buffer; At the same time, the learning rate of the parameter update is adjusted to improve the speed and ability of the model to adapt to data changes.

8. A system for real-time calibration of edge-side weather prediction models, characterized in that, It includes: An initial prediction value output module configured to obtain multi-dimensional meteorological time series data, input into a compressed and optimized lightweight hybrid neural network model, extract spatio-temporal features and output initial prediction values; A calibrated prediction value output module configured to decompose the weight matrix in the lightweight hybrid neural network model into amplitude components and direction components, update the direction components by low-rank matrix, obtain updated direction components, calculate calibration weights based on the amplitude components and the updated direction components, combine local observation data to correct the bias of the initial prediction values, and generate calibrated prediction values; A low-rank matrix parameter update module configured to dynamically manage historical samples and current samples based on a prediction uncertainty index, construct an anti-forgetting training target by combining a weight regularization constraint, and update low-rank matrix parameters to cope with data distribution changes; A quantization processing module configured to quantize the lightweight hybrid neural network model and the low-rank matrix respectively to reduce the consumption of computing and storage resources; A prediction error statistics module configured to monitor prediction error statistics, dynamically adjust the sample storage size and the parameter update rate based on the error changes to maintain stable calibration performance.

9. An electronic device, comprising: It includes: A processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores a computer program executable by an electronic device, and when the computer program runs on the electronic device, the electronic device executes the steps of the method in any one of claims 1 to 7.

Images