A real-time calibration method and system of an edge side weather prediction model, an electronic device and a storage medium

Abstract

The application 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, which comprise the following steps: acquiring multi-dimensional weather time series data, inputting the multi-dimensional weather time series data into a neural network model, extracting space-time features and outputting initial prediction values; decomposing a weight matrix in the neural network model into an amplitude component and a direction component, updating the direction component into a low-rank matrix, obtaining an updated direction component, calculating calibration weights according to the amplitude component and the updated direction component, correcting the initial prediction values in combination with local observation data, and generating calibrated prediction values; dynamically managing historical samples and current samples based on a prediction uncertainty index, constructing an anti-forgetting training target in combination with weight regularization constraints, and updating low-rank matrix parameters to cope with data distribution changes; and the application can improve the reliability of real-time weather prediction services.

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 edge-side weather prediction models, electronic equipment and 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 edge-side weather prediction models to solve the above technical problems. SUMMARY

[0005] The purpose of the present application is to provide a real-time calibration method, system, electronic equipment and storage medium for edge-side weather prediction models 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 edge-side weather prediction models, comprising:

[0007] Obtain multi-dimensional weather time series data and input it into a compressed and optimized lightweight hybrid neural network model to extract spatio-temporal features and output initial prediction values;

[0008] 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, a deviation correction is performed on the initial prediction value combined with local observation data, and a calibrated prediction value is generated;

[0009] Based on the prediction uncertainty index, the historical samples and the current samples are dynamically managed, the anti-forgetting training target is constructed combined with the weight regularization constraint, and the low-rank matrix parameters are updated to cope with the change of data distribution;

[0010] The lightweight hybrid neural network model and the low-rank matrix are quantized respectively to reduce the consumption of calculation and storage resources;

[0011] 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.

[0012] 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:

[0013] The multi-dimensional meteorological time series data is received, including multiple meteorological element observation values at multiple consecutive time steps;

[0014] The input multi-dimensional meteorological time series data is convolved by a one-dimensional convolution layer, multiple local spatio-temporal feature patterns at different time scales are extracted using multiple convolution kernels of different sizes, and a feature sequence containing rich feature information is output;

[0015] The obtained feature sequence is input into a long short-term memory network to capture long-term dependence and temporal dynamic change rules;

[0016] The initial prediction value is output, including prediction results of multiple meteorological elements at one or more future time steps.

[0017] In some embodiments, the lightweight hybrid neural network model is compressed and optimized, further comprising:

[0018] The trained lightweight hybrid neural network model is analyzed for weights, and connections with weights below a preset threshold are identified and removed;

[0019] The remaining weights of the pruned model are quantized with low bits, and the original precision floating-point weight values are mapped to an integer representation range;

[0020] The compressed and optimized lightweight hybrid neural network model is obtained.

[0021] 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, and the updated direction component is calculated based on the magnitude component and the updated direction component.

[0022] Each weight matrix in the lightweight hybrid neural network model is decomposed into two components, a magnitude component representing the weight magnitude and a direction component representing the weight direction information.

[0023] The direction component obtained by decomposition is updated by a low-rank matrix, and the direction component is adjusted and optimized by introducing the product of two low-rank matrices.

[0024] The calibration weight is calculated based on the magnitude component and the updated direction component.

[0025] Based on the calibration weight, the initial prediction value and the local observation data, the calibrated prediction value is calculated by forward propagation.

[0026] In some embodiments, the direction component is updated by a low-rank matrix, further comprising:

[0027] The two low-rank matrices used to update the direction component are quantized by a low-bit, reducing the numerical precision.

[0028] Each element value in the low-rank matrix is mapped from the original precision floating point representation to the integer representation interval.

[0029] In some embodiments, the historical samples and current samples are dynamically managed based on the prediction uncertainty index, and the anti-forgetting training target is constructed by combining the weight regularization constraint, and the low-rank matrix parameters are updated to cope with the data distribution changes, further comprising:

[0030] A sample buffer is established and maintained to store historical observation data and corresponding timestamp information.

[0031] According to the prediction uncertainty index of the newly arrived sample, the sample content in the sample buffer is updated by an uncertainty priority strategy.

[0032] 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.

[0033] 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.

[0034] In some embodiments, 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, further comprising:

[0035] The error statistics between the calibrated prediction values and the actual observation values are continuously calculated and recorded;

[0036] The error statistics are periodically determined whether to exceed the preset threshold, and the preset threshold is determined based on the historical error data distribution characteristics;

[0037] When the error statistics exceed the preset threshold, the sample storage size is dynamically adjusted, and the capacity of the sample buffer is increased;

[0038] The learning rate of the parameter update is adjusted at the same time, and the speed and ability of the model to adapt to data changes are improved.

[0039] Based on the same concept, the present application also provides a real-time calibration system for edge side weather prediction model, comprising:

[0040] An initial prediction value output module is configured to obtain multi-dimensional weather time series data, input into a compressed and optimized lightweight hybrid neural network model, extract spatio-temporal features and output initial prediction values;

[0041] A calibrated prediction value output module is configured to decompose the weight matrix in the lightweight hybrid neural network model into amplitude components and direction components, update the direction components to obtain updated direction components, calculate calibration weights according to the amplitude components and the updated direction components, combine local observation data to correct the deviation of the initial prediction values, and generate calibrated prediction values;

[0042] A low-rank matrix parameter update module is configured to dynamically manage historical samples and current samples based on prediction uncertainty indicators, construct an anti-forgetting training target combined with weight regularization constraints, and update low-rank matrix parameters to cope with data distribution changes;

[0043] A quantization processing module is configured to quantize the lightweight hybrid neural network model and the low-rank matrix respectively, and reduce the consumption of computing and storage resources;

[0044] A prediction error statistics module is configured to 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.

[0045] 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 complete communication 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 real-time calibration method of the edge-side weather prediction model.

[0046] 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 the real-time calibration method of the edge-side weather prediction model.

[0047] Compared with the prior art, the application has the beneficial effects that:

[0048] The application discloses a real-time calibration method and system of an edge-side weather prediction model, an electronic device and a storage medium, which improves parameter efficiency and resource adaptability, and realizes effective calibration by updating only a parameter matrix of extremely low rank through a parameter efficient fine-tuning framework and a rank adaptation technology.

[0049] The application realizes millisecond-level real-time calibration and guarantees prediction timeliness, and since the amount of parameters to be updated is extremely small and the parameters are quantized at a low bit, the calculation cost of the calibration process is extremely low, and the time consumption of single sample updating can be controlled at a millisecond level.

[0050] The application enhances the long-term robustness of the model in a dynamic environment, intelligently filters and memorizes high-value historical data through a continuous learning closed loop that integrates uncertainty management, including an uncertainty-based sample buffer dynamic updating mechanism and an anti-forgetting training strategy, effectively alleviates the problem of catastrophic forgetting, and combines real-time monitoring and drift detection mechanisms of prediction errors to automatically perceive changes in data distribution and dynamically adjust learning strategies, ensuring the stability and accuracy of the calibration system in long-term operation.

[0051] The application takes into account performance and power consumption, facilitates hardware acceleration deployment, and the lightweight model structure and low-precision quantization calculation are very suitable for the architecture characteristics of edge neural processing units (NPUs), facilitating the use of hardware acceleration units to further improve calculation efficiency, achieving inference speed improvement while controlling peak memory occupancy at a very low level, and providing a basis for efficient deployment and application on various edge chip platforms. BRIEF DESCRIPTION OF DRAWINGS

[0052] Other features, objects, and advantages of the application will become more apparent from the following detailed description of non-limiting embodiments, when read in connection with the following accompanying drawings:

[0053] Figure 1 is a flowchart of a real-time calibration method of an edge-side weather prediction model according to some embodiments of the present application;

[0054] Figure 2 is a structural diagram of a real-time calibration system of an edge-side weather prediction model according to some embodiments of the present application;

[0055] Figure 3 is a structural diagram of an electronic device according to some embodiments of the present application;

[0056] In the drawings, 710 is a processor; 720 is a memory; 730 is an input device; and 740 is an output device. DETAILED DESCRIPTION

[0057] In order to make the purposes, technical solutions and advantages of the present application clearer, the following will further describe the present application with reference to the drawings. Obviously, the described embodiments are only some of the embodiments of the present application, but not all of the embodiments of the present application. 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.

[0058] The terms used in the embodiments of the present application are only for the purpose of describing particular embodiments and are not intended to limit the present application. The singular forms "a", "an" and "the" used in the embodiments of the present application and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise. "Plural" generally includes at least two.

[0059] It should be understood that the term "and / or" used herein only describes an 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 a "or" relationship.

[0060] 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.

[0061] 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, 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 [the stated condition or event] is detected" or "in response to detecting [the stated condition or event]."

[0062] It is also important to note that the terms "comprises", "comprising", or other variations thereof are intended to cover a non-exclusive inclusion, such that a product or method that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such product or method. An element proceeded by "comprises... a" does not, without more constraints, exclude the presence of additional identical elements in the product or method that comprises the element.

[0063] In particular, it is to be noted that the symbols and / or numbers present in the description, if not marked in the description of the figures, are not figure references.

[0064] With reference to Figure 1 A real-time calibration method of an edge side weather prediction model, comprising:

[0065] S101, acquire multi-dimensional weather time series data, input into a compressed and optimized lightweight hybrid neural network model, extract spatio-temporal features and output initial prediction values;

[0066] S102, decompose the weight matrix in the lightweight 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, combine local observation data to correct the bias of the initial prediction values, and generate calibrated prediction values;

[0067] S103, dynamically manage historical samples and current samples based on prediction uncertainty indicators, construct an anti-forgetting training target combined with weight regularization constraints, and update low-rank matrix parameters to cope with data distribution changes;

[0068] S104, quantize the lightweight hybrid neural network model and the low-rank matrix respectively to reduce the consumption of calculation and storage resources;

[0069] S105, monitor prediction error statistics, dynamically adjust the sample storage size and parameter update rate based on error changes to maintain stable calibration performance.

[0070] 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 compressed and optimized lightweight hybrid neural network model; the model uses one-dimensional convolutional neural network to extract the local space-time features of the input data, and uses long short-term memory network to model the time series of the features, and outputs the initial meteorological prediction value. To realize efficient parameter updating, the weight matrix in the above neural network is decomposed into a scalar component representing the overall amplitude and a unit matrix component representing the direction information; only the direction component is updated by a low-rank matrix, and the adjustment of the original direction component is realized by introducing the product of two low-rank matrices, and then the calibrated weight is calculated by combining the amplitude component; the initial prediction value is calculated by using the weight, and combined with the real-time collected local observation data, the prediction error is calculated by the loss function, and the calibrated prediction result is generated. To deal with the problem of changing data distribution over time, a sample management strategy based on prediction uncertainty is adopted to dynamically maintain the 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 the weight change regularization constraint is introduced, and the anti-forgetting training objective function is constructed together to maintain the model's memory of past knowledge. To further reduce resource consumption, the weight parameters and low-rank update matrices in the lightweight hybrid neural network model are quantized, 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 error is continuously monitored, and when the error mean or variance exceeds the set threshold, the capacity of the sample buffer and the learning rate in the parameter update process are automatically adjusted, so as to realize the rapid adaptation to environmental changes and maintain the long-term stability of the calibration system.

[0071] 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 lightweight CNN-LSTM model after pruning and 8-bit quantization. The model contains 10 one-dimensional convolution kernels and LSTM units with a hidden layer size of 20, and outputs the initial prediction value of temperature and humidity 1 hour in the future, assuming that the initial prediction is temperature 25°C and humidity 70%. Decompose a certain weight matrix in the model to obtain 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 obtain the new weight. Using the new weight and the actual measured temperature of 24°C and humidity of 72% from the local sensor, the bias loss is calculated to obtain the calibrated prediction value, for example, temperature 24.5°C and humidity 71.5%. The sample buffer capacity is set to 50, and 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 continuously risen by more than 5% of the threshold, temporarily expand the buffer capacity to 100, and at the same time increase the learning rate by 50% to accelerate the model to adapt to the new data pattern, thereby maintaining the prediction accuracy.

[0072] 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, including 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; outputting the initial prediction value, including the prediction result of multiple meteorological elements at one or more future time steps.

[0073] Understandably, the system receives multidimensional meteorological time-series data from sensors. This data includes observations of various meteorological elements collected at multiple consecutive time steps, forming a complete time-series input. This time-series data is then input into a compressed and optimized lightweight hybrid neural network model, which integrates a one-dimensional convolutional layer and a long short-term memory (LSM) network. The one-dimensional convolutional layer performs convolution operations on the input time-series data, using multiple convolutional kernels with different time widths to extract local spatiotemporal feature patterns at different time scales, thereby outputting a feature sequence containing multi-scale feature information. This feature sequence is then input into the LSM network, whose gating mechanism captures long-term dependencies and temporal dynamic changes in the meteorological data. The LSM network outputs an initial prediction value, which includes predictions of various meteorological elements at one or more future time steps, providing a basis for subsequent calibration operations.

[0074] For example, an edge computing node receives temperature, humidity, and air pressure data collected hourly over the past 24 hours, forming an input matrix with 24 time steps and a feature dimension of 3. This matrix is ​​then fed into a lightweight hybrid neural network that has undergone pruning and 8-bit quantization compression. The model's one-dimensional convolutional layer uses three kernels of sizes 2, 3, and 4 to extract short- to medium-term local features, outputting a feature sequence containing 22 time steps. This feature sequence is then fed into an LSTM network with 16 hidden units. The LSTM captures long-term dependencies exceeding 12 hours through its internal control states, outputting predicted temperature and humidity values ​​for the next 3 hours. For example, the predicted values ​​for the first hour are 25.2℃ and 68% humidity, the second hour is 24.8℃ and 71% humidity, and the third hour is 24.5℃ and 73% humidity.

[0075] In some applications, the lightweight hybrid neural network model is compressed and optimized, including performing weight analysis on the trained lightweight hybrid neural network model, identifying and removing connections whose absolute weight values ​​are lower than a preset threshold; performing low-bit quantization on the remaining weights of the pruned model, mapping the original precision floating-point weight values ​​to the integer representation range; and obtaining the compressed and optimized lightweight hybrid neural network model.

[0076] Understandably, weight analysis is performed on a trained lightweight hybrid neural network model. This involves scanning all weight connections in the network and calculating their absolute values ​​to identify connections whose absolute weight values ​​are below a preset threshold. Model pruning is then performed to remove these low-importance connections, reducing the number of parameters and computational complexity. Next, the remaining weights in the pruned model undergo low-bit quantization, converting the original high-precision floating-point weight values ​​to a low-bit-width integer representation using a linear mapping function, thus reducing the storage space requirement for individual weight values. This results in a compressed and optimized lightweight hybrid neural network model that maintains its original predictive performance while reducing computational resource consumption and storage space requirements, enabling deployment on resource-constrained edge computing devices.

[0077] For example, a lightweight hybrid neural network containing convolutional and LSTM layers is compressed and optimized. All weights in the model are analyzed, and an absolute threshold of 0.05 is set. All weights with absolute values ​​below this threshold are removed, reducing the model parameter size from 100KB to 60KB. The remaining weights after pruning are then quantized to 8 bits, mapping the original 32-bit floating-point weight values ​​to an integer range of 0 to 255 through a linear transformation. For example, a floating-point weight value of -0.34 is quantized to the integer 45, and 0.78 is quantized to 200. After compression optimization, the model size is reduced to 25KB, and peak memory usage is reduced to below 15MB, improving deployment efficiency on edge devices.

[0078] In some applications, the weight matrix in the lightweight hybrid neural network model is decomposed into amplitude and direction components. The direction component is updated using a low-rank matrix to obtain the updated direction component. Calibration weights are calculated based on the amplitude and updated direction components. The initial predicted value is then corrected for deviation using local observation data to generate a calibrated predicted value. This process includes decomposing each weight matrix in the lightweight hybrid neural network model into two components: an amplitude component representing the magnitude of the weight and a direction component representing the direction information of the weight. The decomposed direction component is updated using a low-rank matrix, and the direction component is adjusted and optimized by introducing the product of the two low-rank matrices. Calibration weights are calculated based on the amplitude and updated direction components. The calibrated predicted value is then calculated through forward propagation based on the calibration weights, the initial predicted value, and local observation data.

[0079] Understandably, each weight matrix in a lightweight hybrid neural network model is decomposed into two components: an amplitude component representing the overall magnitude of the weights and a unitized directional component representing the direction information of the weights. The resulting directional component is updated using a low-rank matrix. This is achieved by introducing two matrices of a specific rank and calculating their product, then adding this product to the original directional component to adjust and optimize it, resulting in the updated directional component. Matrix multiplication is then performed using the unchanged amplitude component and the updated directional component to calculate the calibration weights. Based on the newly calculated calibration weights, the initial predicted values ​​output by the model, and real-time acquired local observation data, forward propagation is used to calculate more accurate calibrated predicted values, thus correcting the deviation of the initial prediction results.

[0080] For example, a 100×50 weight matrix in the model is decomposed, yielding an amplitude component of 2.5 and a 100×50 identity matrix for the directional component. A low-rank update of rank 4 is performed on the directional component, introducing two low-rank matrices of 100×4 and 4×50 dimensions respectively. Their product is calculated and superimposed on the original directional component. A calibration weight matrix is ​​then calculated based on the amplitude component of 2.5 and the updated directional component. Using the new weights, the initial predicted values ​​of temperature 25.2℃ and humidity 68% are forward-computed. Combined with the locally measured temperature of 24.5℃ and humidity of 72%, and by minimizing the mean square error loss, a calibrated predicted value of temperature 24.8℃ and humidity 70.5% is generated.

[0081] In some applications, updating the direction component using a low-rank matrix further includes low-bit quantization of the two low-rank matrices used to update the direction component, reducing the precision of the numerical representation; and mapping each element value in the low-rank matrix from its original precision floating-point representation to an integer representation range.

[0082] Understandably, during the low-rank matrix update process, the two low-rank matrices used to update the direction components undergo further low-bit quantization processing. This reduces computational resource consumption by decreasing the numerical representation precision of matrix elements. Each element value in the low-rank matrix is ​​transformed from its original high-precision floating-point representation to a low-bit-width integer representation range using a linear mapping function. This reduces the matrix's storage space requirements and memory consumption during computation, while maintaining the matrix's numerical distribution characteristics. This ensures the effectiveness and stability of the low-rank update operation, enabling efficient optimization and adjustment of the direction components even in resource-constrained environments.

[0083] For example, 4-bit quantization is performed on two low-rank matrices used to update direction components. This maps each element value in the first (100×4) and the second (4×50) low-rank matrices from its original 32-bit floating-point representation to an integer range of 0 to 15. Specifically, the floating-point value -0.32 is converted to the integer 3, 0.75 to 12, and 1.28 to 15 using the quantization formula. After 4-bit quantization, the storage space for each matrix element is reduced from 32 bits to 4 bits, and the total storage requirement for the two low-rank matrices is reduced from 18,000 bits to 2,250 bits, a reduction of 87.5% in storage overhead, thus improving computational efficiency on edge devices.

[0084] In some applications, historical and current samples are dynamically managed based on prediction uncertainty indicators. An anti-forgetting training objective is constructed using weighted regularization constraints, and low-rank matrix parameters are updated to cope with changes in data distribution. This includes establishing and maintaining a sample buffer to store historical observation data and corresponding timestamp information; updating the sample content in the sample buffer using an uncertainty-first strategy based on the prediction uncertainty indicators of newly arrived samples; calculating the comparison between the prediction uncertainty of newly arrived samples and the uncertainty of existing samples in the sample buffer in real time to decide whether to store the new sample in the sample buffer; and combining the newly arrived sample data with the historical sample data stored in the sample buffer, and adding weighted regularization constraints, to jointly construct an anti-forgetting training objective function used to update the low-rank matrix parameters.

[0085] Understandably, a sample buffer is established and maintained to store historical observation data and their corresponding timestamps, forming a valuable historical dataset. Based on the prediction uncertainty index of newly arriving samples, an uncertainty-first strategy is adopted to update the sample content in the sample buffer, prioritizing the retention of sample data with higher prediction uncertainty. The prediction uncertainty value of newly arriving samples is calculated in real time and compared with the uncertainty index of existing samples in the sample buffer. The decision on whether to store the new sample in the sample buffer is based on the comparison results. Combining the newly arriving sample data with the historical sample data stored in the sample buffer, and adding weighted regularization constraints, an anti-forgetting training objective function is constructed. Optimizing this objective function updates the parameters of the low-rank matrix, enabling the model to effectively cope with changes in data distribution while avoiding forgetting existing knowledge.

[0086] For example, a sample buffer with a capacity of 100 is established to store temperature and humidity observation data collected over the past 200 hours, along with their corresponding timestamps. The prediction uncertainty of new samples is obtained by calculating the prediction variance. The uncertainty of the current new sample is 0.85, while the sample with the lowest uncertainty in the buffer has a value of 0.35. Since the new sample has higher uncertainty, an uncertainty-first strategy is adopted to add it to the buffer, while samples with an uncertainty of 0.35 are removed. During model updates, 20 historical samples are randomly selected from the buffer and used together with the current new sample to calculate the mean squared error loss. A weight change L2 regularization term is added, with a regularization coefficient set to 0.01, which together constitute the anti-forgetting training objective function. The parameters of the low-rank matrix with a rank of 4 are updated using the gradient descent algorithm, with a learning rate set to 0.001.

[0087] 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 the error statistics between the calibrated predicted values ​​and the actual observed values; periodically determining whether the error statistics exceed a preset threshold, which is determined based on the distribution characteristics of historical error data; when the error statistics exceed the preset threshold, the sample storage size is dynamically adjusted to increase the capacity of the sample buffer; and the learning rate of parameter updates is adjusted to improve the speed and ability of the model to adapt to data changes.

[0088] Understandably, it involves continuously calculating and recording the statistical characteristics of the error between calibrated predicted values ​​and actual observed values, including the mean, variance, and trends of the error over time. Periodically, it's determined whether the statistical characteristics exceed a preset threshold determined based on the distribution characteristics of historical error data; this threshold reflects the range of error fluctuations. When the statistical characteristics exceed the preset threshold, it indicates a potential change in data distribution or a decline in model performance. At this point, the sample storage size is dynamically adjusted, increasing the sample buffer capacity to accommodate more diverse data samples. Simultaneously, the learning rate during parameter updates is adjusted accordingly to improve the model's speed of adaptation to data changes and its learning ability, ensuring long-term stable predictive performance in the face of environmental changes.

[0089] For example, the error between predicted and observed temperatures is continuously calculated. The average error over the past 24 hours was 0.5℃, with a variance of 0.2. When the average error rises to 1.2℃ and exceeds the preset threshold of 0.8℃, an anomaly is identified. The sample buffer capacity is dynamically adjusted from 100 to 150 samples to retain more historical data features. Simultaneously, 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 drops back to 0.6℃, returning to a stable state. At this point, the buffer capacity is adjusted back to 120 samples, and the learning rate is restored to 0.0012 to maintain optimal performance.

[0090] The following describes another embodiment of the real-time calibration method for a marginal meteorological prediction model according to the present invention:

[0091] In this embodiment, the model structure includes:

[0092] The input layer receives multi-dimensional meteorological time-series data (such as temperature, humidity, air pressure, etc., denoted as a sequence). ;

[0093] Input layer parameters:

[0094] :

[0095] This represents the input multi-dimensional meteorological time-series data sequence. Among them, It is the number of time steps (such as the number of hours or days observed). It is the first The feature vectors at each time step.

[0096] :

[0097] Specific observation data for each time step It refers to the number of characteristic dimensions (such as the number of meteorological indicators like temperature, humidity, and air pressure). The larger the value, the more comprehensive the monitored indicators. And the more comprehensive the indicators, the more complex the data is in terms of "features".

[0098] CNN feature extraction layer: through 1 convolution kernel ( Extracting local spatiotemporal features (within the convolution window size), and outputting... ,in:

[0099] ( For activation function, (for bias);

[0100] CNN feature extraction layer parameters:

[0101] :

[0102] The number of convolutional kernels (i.e., the number of feature channels extracted). The larger the value, the more diverse the local patterns it can capture, but the computational cost will also increase.

[0103] :

[0104] Convolution window size (the length of the sliding window in the time dimension). For example... =3 indicates that each convolution operation covers meteorological data for 3 consecutive time steps, which is used to capture short-term dependencies (such as temperature fluctuation features within 3 hours).

[0105] :

[0106] Convolutional kernel weight matrix. Dimensions are... (Number of convolution kernels) × (Window size) × (Input feature dimension), each convolutional kernel is responsible for extracting a local feature pattern (such as the association pattern of "high temperature - humidity decrease").

[0107] :

[0108] The bias term in the convolution operation is used to adjust the baseline of the convolution output, increasing the flexibility of the model.

[0109] :

[0110] Activation functions (such as ReLU and sigmoid) are used to introduce nonlinear transformations, enabling the model to capture complex feature relationships (such as the nonlinear relationship between temperature and air pressure).

[0111] :

[0112] The output feature sequence of the CNN layer. Due to the sliding of the convolutional window, the number of time steps in the output is... (shorter than the input sequence) (Time step).

[0113] :

[0114] The first in the feature sequence The feature vector at each time step is calculated by the convolution kernel and a local window of the input data (formula is given). This includes key local features within that time window.

[0115] LSTM temporal prediction layer: This layer converts feature sequences into... Inputting the data into an LSTM unit, the system uses a gating mechanism to capture long-term dependencies and outputs initial predictions. ( (For the prediction dimension). The LSTM state update formula is:

[0116] ;

[0117] ( These are the input gate, forget gate, and output gate, respectively. In cellular state, In hidden state, This is the weight matrix. (for bias).

[0118] Temporal prediction layer parameters:

[0119] :

[0120] The output of the LSTM layer, i.e., the initial prediction. It is a predictive dimension (e.g., if you need to predict the temperature and humidity for the next hour, then...). =2).

[0121] LSTM state update formula parameters:

[0122] LSTM dynamically adjusts the information flow through input gates, forget gates, and output gates. Its core parameters are as follows:

[0123] :

[0124] The first The input gate, forget gate, and output gate (all vectors) of the time step. Their values ​​are in the range (0, 1) (activated by sigmoid), controlling the "inflow," "retention," and "outflow" of information.

[0125] Input gate : Determine the current input features And the previous hidden state In this context, what information needs to be stored in the cell state?

[0126] Forgotten Gate : Determines the cell state at the previous moment In the process, which information needs to be retained (such as whether the abnormally high temperature a few days ago was "forgotten").

[0127] Output gate Determines the current cell state In the hidden state, which information needs to be output? (This will affect the final prediction).

[0128] :

[0129] : No. The cell state at a time step ("memory unit") stores long-term information and is composed of the historical state filtered by the forget gate and the new information filtered by the input gate. ).

[0130] : Candidate cell state, calculated from the current input and the previous hidden state ( ), which contains new candidate information.

[0131] :

[0132] No. The hidden state at the time step is obtained by tanh activation of the cell state controlled by the output gate. ), is the instantaneous output of the LSTM, used for subsequent predictions or passed to the next time step.

[0133] :

[0134] These are the weight matrices for the input gate, forget gate, output gate, and candidate cell states, respectively, used to weight the input features. And the previous hidden state Mapped to a gating signal or candidate state.

[0135] : The corresponding bias term is used to adjust the baseline of the gating signal.

[0136] :

[0137] Element-wise multiplication (Hadamard product) is used to filter or superimpose information element by element (such as the multiplication of the forget gate with the historical cell state to achieve "selective forgetting").

[0138] Optimization strategy: Prune the model (remove weights whose absolute value is less than a threshold). (connection) and 8-bit quantization (to connect weights) Mapped to Interval:

[0139] This compresses the model size to a level that edge devices can handle (typically <100MB, depending on the specific situation).

[0140] Second layer: DoRA real-time calibration module

[0141] Decoupled Rank Adaptation (DoRA) technique is employed as the core for efficient parameter fine-tuning. While freezing the core model weights, localized bias correction is achieved through low-rank updates, balancing calibration accuracy and resource efficiency. Weight decomposition principle:

[0142] Arbitrary weight matrix in the core model It is decomposed into amplitude and direction components: ,in (F-norm standardized amplitude)

[0143] (Unit direction matrix).

[0144] Note: F-norm normalized magnitude is a normalization operation that first calculates the F-norm of a matrix (or vector) (the square root of the sum of the squares of its elements), and then divides each element of the original matrix (or vector) by the F-norm, thereby scaling its magnitude to the unit F-norm.

[0145] DoRA only performs low-rank updates on the direction components, avoiding the weight norm drift problem in LoRA:

[0146] ( It is a low-rank matrix. (The rank is usually 4-8).

[0147] The final calibrated weights are .

[0148] Calibration process:

[0149] Input core model output Compared with local observations The bias correction function is learned through the DoRA module. Output the calibrated predicted value:

[0150] ;

[0151] The formula is explained in detail below:

[0152] Symbol meaning:

[0153] : The final predicted value after calibration (such as calibrated temperature, humidity and other meteorological indicators).

[0154] : The initial predicted value output by the core model (CNN+LSTM hybrid architecture).

[0155] : The bias correction function learned by the DoRA module is used to correct the prediction bias of the core model.

[0156] Trainable parameters of the DoRA module (low-rank matrix) and It dynamically adjusts by learning local observation data in real time.

[0157] Formula logic:

[0158] The calibrated prediction result = initial prediction value of the core model + bias correction amount calculated by the DoRA module, where the bias correction function... Initial predictions from the core model As input, low-rank parameters based on DoRA The deviation value that needs to be corrected is calculated, and finally the dynamic calibration of the initial prediction is achieved.

[0159] Core function:

[0160] While freezing the weights of the core model, localized bias correction is achieved with a very small number of parameters (only <1% of the core model), taking into account the resource constraints of edge devices and the prediction accuracy requirements, and supporting real-time updates (single sample update time <10ms).

[0161] Update method:

[0162] By minimizing the loss function Updated in real time.

[0163] Quantization adaptation: Combining 4-bit quantization (QLoRA variant) to transform low-rank matrices Quantization to the [0,15] interval using the zero-pole quantization formula ( Zero point (The scaling factor) further reduces memory usage and ensures real-time training on edge devices (single sample update time < 10ms).

[0164] The process is explained as follows:

[0165] Core objective:

[0166] By freezing the core model weights and using only low-rank matrices and Based on real-time calibration of the trainable parameters of the DoRA module, further compression is performed. and Minimize memory usage 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.

[0167] 4-bit quantization (QLoRA variant) means: quantization transforms matrix elements that would otherwise be stored as 32-bit or 16-bit floating-point numbers into quantized values. and The numerical value in the text is compressed into a 4-bit integer (i.e., (a number of possible values), thereby reducing memory usage to its original level. (32-bit → 4-bit) or (16 bits → 4 bits).

[0168] Quantization range: and The element values ​​are mapped to the range [0, 15] (the range of 4-bit integers), replacing the original floating-point representation.

[0169] Zero-pole quantization:

[0170] : Raw floating-point numbers (low-rank matrix) or (a certain element in).

[0171] Offset is a constant to be determined and used to adjust the baseline of the quantization range (e.g., mapping negative numbers to the non-negative interval [0, 15]).

[0172] The scaling factor (scale) controls the mapping ratio from the original value to the quantization interval. The calculation formula is typically: (The dynamic range of the original value is uniformly mapped to 16 discrete values).

[0173] Rounding ensures the result is an integer.

[0174] Calculation example:

[0175] If the original value The range is ,but (Midnight) (Scaling factor). For The quantized value is (falling) Within the interval).

[0176] Functions and advantages:

[0177] 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.

[0178] 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.

[0179] 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.

[0180] Third layer: Continuous learning loop

[0181] 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.

[0182] Rehearsal buffer design:

[0183] 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.

[0184] Anti-forgetting training:

[0185] 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:

[0186] ;

[0187] ( For the loss of new samples, For buffer sample loss, For hyperparameters, (Parameters from the previous round);

[0188] Drift detection: Calculate prediction error using a sliding window. mean With variance ,when ( When the initial error mean is reached, the buffer is expanded (temporarily increased to) ) and learning rate improvement ( This will accelerate adaptation to sudden changes in data distribution.

[0189] Efficiency can be improved by leveraging neural network processors (NPUs) integrated into edge devices in the following ways:

[0190] Low-precision computation: Mapping CNN convolution and LSTM gating operations to the 8-bit arithmetic unit of the NPU improves inference speed by 3-5 times;

[0191] Memory reuse: By sharing weights between model layers (such as the gated weight sharing matrix of LSTM), the peak memory usage is controlled to <32MB.

[0192] Software toolchain:

[0193] Model training: The core model is trained on the server side using PyTorch, and the DoRA module is initialized using the peft library;

[0194] Device deployment: The model was converted to TFLite format using TensorFlow Lite for Microcontrollers, and the inference and training logic was implemented in C++;

[0195] Monitoring and Updates: The edge device periodically uploads calibration error statistics to the cloud. The cloud aggregates and updates the DoRA initialization parameters through federated learning and pushes them to the device to optimize the initial calibration performance.

[0196] For the purpose of simplicity, the method steps disclosed in the above embodiments are described as a series of actions. However, those skilled in the art should understand that the embodiments of the present invention are not limited to the described order of actions, because according to the embodiments of the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily essential to the embodiments of the present invention.

[0197] like Figure 2 As shown, the present invention also provides a real-time calibration system for a marginal-side meteorological forecasting model, comprising:

[0198] The initial prediction output module 201 is configured to acquire multidimensional meteorological time series data, input it into a compressed and optimized lightweight hybrid neural network model, extract spatiotemporal features, and output the initial prediction value.

[0199] The calibrated prediction output module 202 is configured to decompose the weight matrix in the lightweight hybrid neural network model into amplitude components and direction components, update the direction components with a low-rank matrix to obtain updated direction components, calculate calibration weights based on the amplitude components and updated direction components, and perform deviation correction on the initial prediction values ​​in combination with local observation data to generate calibrated prediction values.

[0200] The low-rank matrix parameter update module 203 is configured to dynamically manage historical and current samples based on the prediction uncertainty index, construct an anti-forgetting training objective by combining weight regularization constraints, and update the low-rank matrix parameters to cope with changes in data distribution.

[0201] The quantization processing module 204 is configured to perform quantization processing on the lightweight hybrid neural network model and the low-rank matrix respectively, thereby reducing the consumption of computing and storage resources.

[0202] The prediction error statistics module 205 is configured to monitor the prediction error statistics and dynamically adjust the sample storage size and parameter update rate based on error changes in order to maintain stable calibration performance.

[0203] It is worth noting that although only some basic functional modules are disclosed in the embodiments of this invention, it does not mean that the composition of this system is limited to the above-mentioned basic functional modules. On the contrary, what this embodiment intends to express is that, based on the above-mentioned basic functional modules, those skilled in the art can arbitrarily add one or more functional modules in combination with existing technology to form an infinite number of embodiments or technical solutions. That is to say, this system is open rather than closed. The fact that this embodiment only discloses a few basic functional modules should not be considered as the scope of protection of the claims of this invention being limited to the disclosed basic functional modules. At the same time, for the convenience of description, the above device is described separately according to its functions as various units and modules. Of course, in implementing this invention, the functions of each unit and module can be implemented in one or more software and / or hardware.

[0204] like Figure 3 As shown, the present invention also provides an electronic device, including: 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 performs the steps of a real-time calibration method for an edge-side weather prediction model.

[0205] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. For example... Figure 3 The structure shown in this embodiment of the invention includes an electronic device comprising one or more processors 710 and a memory 720; the processors 710 in this electronic device may be one or more. Figure 3 Taking a processor 710 as an example; a memory 720 is used 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 a real-time calibration method for an edge-side meteorological prediction model as described in any one of the embodiments of the present invention.

[0206] The electronic device may also include an input device 730 and an output device 740.

[0207] The processor 710, memory 720, input device 730, and output device 740 in this electronic device can be connected via a bus or other means.Figure 3 Taking the example of a connection between China and Israel via a bus.

[0208] The memory 720 in this electronic device serves as a computer-readable storage medium, capable of storing one or more programs. These programs can be software programs, computer-executable programs, or modules, such as the program instructions / modules corresponding to the real-time calibration method for an edge-side weather forecasting model provided in this embodiment of the invention. The processor 710 executes various functional applications and data processing of the electronic device by running the software programs, instructions, and modules stored in the memory 720, thereby implementing the real-time calibration method for an edge-side weather forecasting model described in the above method embodiment.

[0209] The memory 720 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the electronic device. Furthermore, the memory 720 may include high-speed random access memory and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some instances, the memory 720 may further include memory remotely located relative to the processor 710, which can be connected to the device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0210] Input device 730 can be used to receive input digital or character information, and to generate key signal inputs related to user settings and function control of the electronic device. Output device 740 may include display devices such as a display screen.

[0211] The present invention also provides a computer-readable storage medium storing a computer program executable by an electronic device, which, when run on the electronic device, causes the electronic device to perform the steps of a real-time calibration method for an edge-side weather forecasting model.

[0212] Specifically, the computer storage medium in this embodiment of the invention can be any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. For example, a computer-readable storage medium can be—but is not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this embodiment, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0213] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A real-time calibration method for a marginal-side meteorological forecasting model, characterized in that, include: Acquire multidimensional meteorological time-series data, input it into a compressed and optimized lightweight hybrid neural network model, extract spatiotemporal features, and output initial prediction values; The weight matrix in the lightweight hybrid neural network model is decomposed into magnitude components and direction components. The direction components are updated by low-rank matrix updates to obtain updated direction components. The calibration weights are calculated based on the magnitude components and the updated direction components. The initial prediction values ​​are then corrected for deviations using local observation data to generate calibrated prediction values. Based on the prediction uncertainty index, historical samples and current samples are dynamically managed. Combined with weight regularization constraints, an anti-forgetting training objective is constructed, and the parameters of the low-rank matrix are updated to cope with changes in data distribution. The lightweight hybrid neural network model and the low-rank matrix are quantized to reduce computational and storage resource consumption. Monitor the prediction error statistics and dynamically adjust the sample storage size and parameter update rate based on error changes to maintain stable calibration performance; Among them, weight analysis is performed on the trained lightweight hybrid neural network model to identify and remove connections whose absolute weight value is lower than a preset threshold; The remaining weights of the pruned model are subjected to low-bit quantization to map the original precision floating-point weight values ​​to the integer representation range. The compressed and optimized lightweight hybrid neural network model is obtained; In this lightweight hybrid neural network model, each weight matrix is ​​decomposed into two components: an amplitude component representing the magnitude of the weight and a directional component representing the direction information of the weight. The directional components obtained from the decomposition are updated with low-rank matrices. The directional components are adjusted and optimized by introducing the product of two low-rank matrices. The calibration weights are calculated based on the amplitude components and the updated direction components. Based on the calibration weights, the initial predicted values, and the local observation data, the calibrated predicted values ​​are calculated through forward propagation. Among these measures, a sample buffer is established and maintained to store historical observation data and corresponding timestamp information. Based on the prediction uncertainty index of newly arrived samples, the sample content in the sample buffer is updated using an uncertainty-first strategy. The prediction uncertainty of a newly arrived sample is calculated in real time and compared with the uncertainty of existing samples in the sample buffer to decide whether to store the new sample in the sample buffer. By combining newly arrived sample data with historical sample data stored in the sample buffer and adding weight regularization constraints, an anti-forgetting training objective function is constructed to update the parameters of the low-rank matrix.

2. The real-time calibration method for a marginal-side meteorological prediction model according to claim 1, characterized in that, Acquire multidimensional meteorological time-series data, input it into a compressed and optimized lightweight hybrid neural network model, extract spatiotemporal features, and output initial predicted values, further including: Receive the multidimensional meteorological time-series data, including observations of various meteorological elements at multiple consecutive time steps; One-dimensional convolutional layers are used to perform convolution operations on the input multi-dimensional meteorological time series data. Multiple convolutional kernels of different sizes are used to extract local spatiotemporal feature patterns at various time scales, and the output feature sequence contains rich feature information. The obtained feature sequences are input into a long short-term memory network to capture long-term dependencies and temporal dynamic changes. The initial forecast value is output, including forecasts of various meteorological elements at one or more future time steps.

3. The real-time calibration method for a marginal-side meteorological prediction model according to claim 1, characterized in that, The low-rank matrix update of the directional components also includes: Low-bit quantization is performed on the two low-rank matrices used to update the direction components, which reduces the precision of the numerical representation. Map each element value in the low-rank matrix from its original precision floating-point representation to an integer representation range.

4. The real-time calibration method for a marginal-side meteorological prediction model according to claim 1, characterized in that, Monitoring prediction error statistics and dynamically adjusting sample storage size and parameter update rate based on error changes to maintain stable calibration performance, further including: Continuously calculate and record the statistical characteristics of the error between the calibrated predicted values ​​and the actual observed values; Periodically determine whether the error statistical characteristics exceed a preset threshold, the preset threshold being determined based on the distribution characteristics of historical error data; When the error statistics exceed the preset threshold, the sample storage size is dynamically adjusted and the capacity of the sample buffer is increased. At the same time, adjust the learning rate for parameter updates to improve the model's speed and ability to adapt to data changes.

5. A real-time calibration system for a marginal-side meteorological forecasting model, used in the real-time calibration method for a marginal-side meteorological forecasting model as described in claim 1, characterized in that, include: The initial prediction output module is configured to acquire multidimensional meteorological time series data, input it into a compressed and optimized lightweight hybrid neural network model, extract spatiotemporal features, and output the initial prediction value. The calibrated prediction output module is configured to decompose the weight matrix in the lightweight hybrid neural network model into amplitude components and direction components, update the direction components with a low-rank matrix to obtain updated direction components, calculate calibration weights based on the amplitude components and updated direction components, and perform deviation correction on the initial prediction values ​​in combination with local observation data to generate calibrated prediction values. The low-rank matrix parameter update module is configured to dynamically manage historical and current samples based on the prediction uncertainty index, construct an anti-forgetting training objective by combining weight regularization constraints, and update the low-rank matrix parameters to cope with changes in data distribution. The quantization processing module is configured to perform quantization processing on the lightweight hybrid neural network model and the low-rank matrix respectively, thereby reducing the consumption of computing and storage resources. The prediction error statistics module is configured to monitor prediction error statistics and dynamically adjust the sample storage size and parameter update rate based on error changes to maintain stable calibration performance.

6. An electronic device, characterized in that, include: The system 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 via the communication bus; the memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the method according to any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, It stores a computer program executable by an electronic device, which, when run on the electronic device, causes the electronic device to perform the steps of the method according to any one of claims 1 to 4.

Images