A multi-time scale oriented geomagnetic ap index intelligent prediction method and system

Abstract

The application relates to the technical field of weather forecasting, in particular to a multi-time-scale-oriented geomagnetic Ap index intelligent forecasting method and system, which comprises the following steps: acquiring geomagnetic Ap index historical sequence data and corresponding solar wind parameter real-time monitoring data, and performing time sequence alignment on the geomagnetic Ap index historical sequence data to obtain an initial geomagnetic activity data set; performing feature screening on the solar wind parameters in the initial geomagnetic activity data set, calculating the nonlinear correlation coefficient of the solar wind parameters and the geomagnetic Ap index, and obtaining a key solar wind driving parameter subset. The application can effectively process the prediction requirements of different time scales by combining short-term prediction and medium and long-term prediction and adopting a weighted fusion strategy, can provide more accurate geomagnetic Ap index prediction, and the hierarchical prediction mode is particularly suitable for coping with geomagnetic activities of different time periods.

Description

Technical Field

[0001] This invention relates to the field of meteorological forecasting technology, specifically to an intelligent forecasting method and system for geomagnetic Ap index at multiple time scales. Background Technology

[0002] Currently, traditional prediction models are prone to overfitting on training data, resulting in poor performance on test data. This is especially true when geomagnetic activity is complex and variable, where traditional methods may not be able to effectively distinguish between noise and signal. Furthermore, traditional methods may not be able to extract multi-level frequency components from geomagnetic activity sequences, causing important periodic or trend information to be ignored, thus affecting the accuracy of predictions.

[0003] Furthermore, traditional methods often fail to effectively filter out the features most relevant to the target variable when processing input features, leading to model redundancy and reduced prediction efficiency. Moreover, traditional models usually lack sufficient flexibility to adjust prediction strategies when facing geomagnetic activity at different time scales, resulting in unstable prediction results, especially when the state of geomagnetic activity changes. In addition, traditional methods are based on linear assumptions, while geomagnetic activity usually exhibits complex nonlinear characteristics, and traditional linear models often fail to effectively capture this nonlinear relationship, affecting the accuracy of prediction. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides the following technical solution: a smart forecasting method for geomagnetic Ap index across multiple time scales, comprising:

[0005] Historical sequence data of geomagnetic Ap index and corresponding real-time monitoring data of solar wind parameters were acquired, and the historical sequence data of geomagnetic Ap index were time-series aligned to obtain an initial geomagnetic activity dataset. Feature filtering was performed on the solar wind parameters in the initial geomagnetic activity dataset, and the nonlinear correlation coefficient between solar wind parameters and geomagnetic Ap index was calculated to obtain a subset of key solar wind driving parameters.

[0006] With minimizing signal reconstruction error as the optimization objective, the penalty factor and decomposition level of the variational mode decomposition algorithm are globally optimized based on a subset of key solar wind driving parameters to obtain the optimal variational mode decomposition parameters;

[0007] The geomagnetic Ap exponent historical sequence data is adaptively decomposed using the optimal variational mode decomposition parameters to separate high-frequency noise components and eigenmode functions of different frequencies, thereby obtaining a multi-scale geomagnetic signal sequence.

[0008] Multi-scale geomagnetic signal sequences are input into a long short-term memory network model to learn the nonlinear evolution law in the short term and output a short-term prediction sequence of the geomagnetic Ap index for the next 1 to 3 days. A gated recurrent unit neural network model is constructed, and the hidden layer state of the model is initialized according to the 27-day recurrence cycle characteristics of geomagnetic activity to obtain a medium- and long-term prediction model.

[0009] By using a medium- to long-term prediction model, features are extracted from historical geomagnetic Ap index data to capture the periodic evolution trend caused by high-speed flow in coronal holes, and a medium- to long-term prediction sequence of geomagnetic Ap index for the next few weeks to months is output.

[0010] Based on the forecast timeliness requirements, the short-term and medium-to-long-term forecast sequences of the geomagnetic Ap index are weighted and fused. An adaptive weight allocation strategy is used to coordinate the forecast results at different time scales to obtain the final intelligent forecast result of the geomagnetic Ap index.

[0011] Preferably, the solar wind parameters in the initial geomagnetic activity dataset are feature-selected, and the nonlinear correlation coefficient between the solar wind parameters and the geomagnetic Ap index is calculated to obtain a subset of key solar wind driving parameters, including:

[0012] Calculate the mutual information value and Pearson correlation coefficient between each solar wind parameter and the geomagnetic Ap index in the initial geomagnetic activity dataset;

[0013] The kernel density estimation method is used to perform a nonlinear mapping between mutual information value and Pearson correlation coefficient to obtain a comprehensive correlation evaluation index;

[0014] The solar wind parameters are sorted in descending order based on the comprehensive correlation evaluation index, and the parameters that rank at the top of the preset proportion are selected as key driving factors.

[0015] The selected key driving factors constitute a subset of key solar wind driving parameters.

[0016] Preferably, with minimizing signal reconstruction error as the optimization objective, the penalty factor and decomposition level of the variational mode decomposition algorithm are globally optimized based on a subset of key solar wind driving parameters to obtain the optimal variational mode decomposition parameters, including:

[0017] The range of the penalty factor and the range of the number of decomposition levels for the variational mode decomposition algorithm are defined as the search space;

[0018] Using a subset of key solar wind driving parameters as input, the geomagnetic Ap exponent sequence is reconstructed using the variational mode decomposition algorithm, and the root mean square error between the reconstructed sequence and the original sequence is calculated.

[0019] The root mean square error is used as the fitness function value, and a genetic algorithm is used to perform an iterative search within the search space.

[0020] In each iteration, the combination of penalty factor and decomposition level is updated through selection, crossover, and mutation operations;

[0021] When the change in the fitness function value over multiple generations is less than the preset convergence threshold, the iteration stops, and the current optimal penalty factor and decomposition level are output as the optimal variational mode decomposition parameters.

[0022] Preferably, the geomagnetic Ap exponent historical sequence data is adaptively decomposed using optimal variational mode decomposition parameters to separate high-frequency noise components and intrinsic mode functions of different frequencies, resulting in a multi-scale geomagnetic signal sequence, including:

[0023] A constrained variational problem for constructing a variational mode decomposition model using optimal variational mode decomposition parameters;

[0024] The constrained variational problem is solved by using the alternating direction multiplier method, which decomposes the historical sequence data of the geomagnetic Ap index into several eigenmode functions and a residual component.

[0025] Based on the center frequency of each intrinsic mode function, high-frequency random fluctuation components are identified as noise and eliminated, while low-frequency and mid-frequency components with clear physical meaning are retained.

[0026] The retained intrinsic mode functions are arranged in descending order of frequency to form a multi-scale geomagnetic signal sequence.

[0027] Preferably, the multi-scale geomagnetic signal sequence is input into a long short-term memory network model to learn the nonlinear evolution law in the short term, and outputs a short-term prediction sequence of the geomagnetic Ap index for the next 1 to 3 days, including:

[0028] Construct a long short-term memory network model structure that includes an input layer, a hidden layer, and an output layer, and set the number of neurons and activation function in the hidden layer;

[0029] Using multi-scale geomagnetic signal sequences as training samples, the backpropagation algorithm is employed to update the weights and biases of the long short-term memory network model over time.

[0030] During model training, a dropout mechanism is introduced to prevent overfitting, and an early stopping strategy is used to determine the optimal training rounds.

[0031] By using a trained long short-term memory network model to perform forward inference on the current multi-scale geomagnetic signal sequence, a short-term prediction sequence of the geomagnetic Ap index for the next 1 to 3 days is obtained.

[0032] Preferably, a gated recurrent unit neural network model is constructed, and the hidden layer state of the model is initialized according to the 27-day recurrence cycle characteristics of geomagnetic activity to obtain a medium- and long-term prediction model, including:

[0033] Construct a gated recurrent unit neural network model and set up a memory unit inside the network specifically for storing periodic information;

[0034] Based on the 27-day recurrence period in the historical sequence data of the geomagnetic Ap index, the average period length and phase shift are calculated.

[0035] The initial hidden state of the gated recurrent unit neural network is assigned values ​​using the average period length and phase offset, so that the model has prior knowledge of the next 27-day cycle at the start of prediction.

[0036] The assigned gated recurrent unit neural network model is used as a medium- to long-term prediction model with a periodic memory mechanism.

[0037] Preferably, a medium- to long-term prediction model is used to extract features from historical geomagnetic Ap index data to capture the periodic evolution trend caused by coronal hole high-speed flow, and output a medium- to long-term prediction sequence of geomagnetic Ap index for the next few weeks to months, including:

[0038] Input historical data of geomagnetic Ap index into a medium- to long-term prediction model with a periodic memory mechanism according to the time step.

[0039] The retention and forgetting of historical information are controlled by using update gates and reset gates, with a focus on strengthening the memory weight of the 27-day cycle and its harmonic components;

[0040] The hidden layer state is mapped to the geomagnetic Ap index prediction for the next few weeks to months through a fully connected layer;

[0041] By continuously rolling the forecast of future time steps, a complete medium- and long-term forecast sequence of the geomagnetic Ap index is obtained by splicing the results.

[0042] Preferably, based on the forecast lead time requirement, the short-term and medium-to-long-term forecast sequences of the geomagnetic Ap index are weighted and fused. An adaptive weight allocation strategy is used to coordinate the forecast results at different time scales, resulting in the final intelligent forecast result for the geomagnetic Ap index, including:

[0043] Obtain the estimated time from the current forecast time for the solar wind shock to reach Earth, as well as the probability of a geomagnetic storm occurring;

[0044] Based on the estimated time and the probability of geomagnetic storm occurrence, the confidence weights of the short-term and medium-to-long-term prediction branches are dynamically calculated using fuzzy logic.

[0045] If the period is characterized by short-term disturbances, increase the weighting coefficient of the short-term forecast sequence; if the period is characterized by calm, increase the weighting coefficient of the medium- and long-term forecast sequences.

[0046] The weighted short-term and medium-to-long-term prediction sequences are linearly superimposed to obtain the final intelligent prediction result of the geomagnetic Ap index.

[0047] Preferably, the confidence weights of the short-term and medium-to-long-term prediction branches are dynamically calculated using fuzzy logic, including:

[0048] The fluctuation variance of the geomagnetic Ap index and the rate of change of solar wind speed are used as input variables for fuzzy logic;

[0049] Define a fuzzy set that reflects the uncertainty of prediction, and establish a fuzzy rule table between input variables and confidence weights;

[0050] Based on the current input variable values, the specific weight values ​​are calculated using the defuzzification method.

[0051] The calculated weight values ​​are normalized to ensure that the sum of the short-term forecast weight and the medium- and long-term forecast weight is the same.

[0052] A smart forecasting system for geomagnetic Ap index at multiple time scales, applicable to the aforementioned smart forecasting method for geomagnetic Ap index at multiple time scales, includes:

[0053] The feature filtering module is used to acquire historical sequence data of geomagnetic Ap index and corresponding real-time monitoring data of solar wind parameters, and to perform time-series alignment of the historical sequence data of geomagnetic Ap index to obtain an initial geomagnetic activity dataset; to perform feature filtering on solar wind parameters in the initial geomagnetic activity dataset, to calculate the nonlinear correlation coefficient between solar wind parameters and geomagnetic Ap index, and to obtain a subset of key solar wind driving parameters;

[0054] The mode decomposition module is used to optimize the signal reconstruction error by globally optimizing the penalty factor and the number of decomposition layers of the variational mode decomposition algorithm based on a subset of key solar wind driving parameters, so as to obtain the optimal variational mode decomposition parameters.

[0055] The noise separation module is used to adaptively decompose the historical sequence data of the geomagnetic Ap index using the optimal variational mode decomposition parameters, separate high-frequency noise components and eigenmode functions of different frequencies, and obtain multi-scale geomagnetic signal sequences.

[0056] The index prediction module is used to input multi-scale geomagnetic signal sequences into a long short-term memory network model, learn the nonlinear evolution law in the short term, and output a short-term prediction sequence of the geomagnetic Ap index for the next 1 to 3 days; construct a gated recurrent unit neural network model, and initialize the hidden layer state of the model according to the 27-day recurrence cycle characteristics of geomagnetic activity to obtain a medium- and long-term prediction model.

[0057] The feature extraction module is used to extract features from historical geomagnetic Ap index data using a medium- to long-term prediction model, capture the periodic evolution trend caused by high-speed flow in coronal holes, and output medium- to long-term prediction sequences of geomagnetic Ap index for the next few weeks to months.

[0058] The weighted fusion module is used to perform weighted fusion of short-term and medium-to-long-term prediction sequences of the geomagnetic Ap index according to the forecast timeliness requirements. It uses an adaptive weight allocation strategy to coordinate the prediction results at different time scales and obtain the final intelligent forecast result of the geomagnetic Ap index.

[0059] Compared with the prior art, the beneficial effects of the present invention are:

[0060] (1) By combining short-term and medium-to-long-term forecasts and adopting a weighted fusion strategy, this invention can effectively handle the forecasting needs of different time scales and provide more accurate geomagnetic Ap index forecasts. This hierarchical forecasting method is particularly suitable for dealing with geomagnetic activities of different time periods, such as daily short-term disturbances and long-term solar wind change cycles. Moreover, by using variational mode decomposition algorithm and gated recurrent unit neural network model, it can extract multi-level frequency components from geomagnetic activity sequences, separate noise and signals, and avoid overfitting or information loss problems in traditional methods. This improves the robustness and accuracy of the model.

[0061] (2) This invention selects solar wind parameters that are strongly correlated with the geomagnetic Ap index by nonlinear correlation coefficient calculation and kernel density estimation method, which effectively reduces the redundancy of input features, enabling the model to focus on the driving factors with greater influence, thereby enhancing the accuracy and efficiency of prediction. Moreover, by dynamically adjusting the weights of short-term and medium-to-long-term prediction results through fuzzy logic method, the prediction model can automatically adapt the prediction results according to the current geomagnetic activity state, further improving the adaptability and accuracy of the model. Attached Figure Description

[0062] Figure 1 This is a schematic flowchart of the overall method in one embodiment of the present invention;

[0063] Figure 2 This is a schematic diagram of the overall system architecture in one embodiment of the present invention.

[0064] In the diagram: 1. Feature selection module; 2. Modality decomposition module; 3. Noise separation module; 4. Exponential prediction module; 5. Feature extraction module; 6. Weighted fusion module. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] Example 1, please refer to Figure 1 This invention provides a technical solution: a smart forecasting method for geomagnetic Ap index across multiple time scales, comprising:

[0067] S1. Obtain historical sequence data of geomagnetic Ap index and corresponding real-time monitoring data of solar wind parameters, and perform time-series alignment on the historical sequence data of geomagnetic Ap index to obtain an initial geomagnetic activity dataset; perform feature filtering on solar wind parameters in the initial geomagnetic activity dataset, calculate the nonlinear correlation coefficient between solar wind parameters and geomagnetic Ap index, and obtain a subset of key solar wind driving parameters.

[0068] S2. With minimizing signal reconstruction error as the optimization objective, the penalty factor and decomposition level of the variational mode decomposition algorithm are globally optimized based on a subset of key solar wind driving parameters to obtain the optimal variational mode decomposition parameters.

[0069] S3. Adaptive decomposition of historical geomagnetic Ap index data is performed using optimal variational mode decomposition parameters to separate high-frequency noise components and intrinsic mode functions of different frequencies, thereby obtaining multi-scale geomagnetic signal sequences.

[0070] S4. Input the multi-scale geomagnetic signal sequence into the long short-term memory network model, learn the nonlinear evolution law in the short term, and output the short-term prediction sequence of the geomagnetic Ap index for the next 1 to 3 days; construct a gated recurrent unit neural network model, and initialize the hidden layer state of the model according to the 27-day recurrence cycle characteristics of geomagnetic activity to obtain the medium- and long-term prediction model.

[0071] S5. Use a medium- to long-term prediction model to extract features from historical geomagnetic Ap index data, capture the periodic evolution trend caused by high-speed flow in coronal holes, and output a medium- to long-term prediction sequence of geomagnetic Ap index for the next few weeks to months.

[0072] S6. Based on the forecast timeliness requirements, the short-term and medium-to-long-term forecast sequences of the geomagnetic Ap index are weighted and fused. An adaptive weight allocation strategy is used to coordinate the forecast results at different time scales to obtain the final intelligent forecast result of the geomagnetic Ap index.

[0073] It should be noted that the historical data sequence of the geomagnetic Ap index and the real-time monitoring data of the related solar wind parameters are obtained. In order to make these data effective for comparison, the historical data of the geomagnetic Ap index and the solar wind data need to be time-aligned to ensure that they are synchronized. This gives us an initial geomagnetic activity dataset.

[0074] The purpose of feature screening of solar wind parameters is to identify which solar wind parameters have the strongest relationship with the geomagnetic Ap index. By calculating the nonlinear correlation between solar wind parameters and the geomagnetic Ap index, a key subset of solar wind driving parameters was identified. These parameters will serve as inputs for subsequent forecasts.

[0075] To ensure the accuracy of data processing, the system uses the variational mode decomposition (VMD) algorithm to decompose historical data of the geomagnetic Ap index. The goal of this step is to optimize the penalty factor and the number of decomposition layers of VMD based on the selected solar wind parameters to ensure that the signal reconstruction error is minimized. By optimizing the VMD parameters, the system can better separate high-frequency noise components and eigenmode functions of different frequencies in the data.

[0076] Under the optimized VMD parameters, historical data of the geomagnetic Ap index are adaptively decomposed to obtain multi-scale geomagnetic signal sequences; this process helps to capture the variation patterns of geomagnetic activity at different time scales in greater detail.

[0077] Multi-scale signal sequences are input into a Long Short-Term Memory (LSTM) network model. The LSTM model can effectively capture the nonlinear evolution of geomagnetic activity in the short term and output the geomagnetic Ap index prediction results for the next 1 to 3 days.

[0078] For medium- and long-term forecasts, the system also constructs a gated recurrent unit (GRU) neural network model. The GRU model initializes the hidden layer state by considering the 27-day periodicity of geomagnetic activity, thereby obtaining a medium- and long-term prediction model that can capture the periodic evolution trend caused by the high-speed flow of coronal holes and output the geomagnetic Ap index prediction results for the next few weeks to months.

[0079] Depending on the different forecast lead time requirements, the system will weight and fuse the results of short-term and medium- to long-term forecasts; through an adaptive weight allocation strategy, the forecast results of these two time scales will be coordinated to finally generate accurate intelligent forecast results of the geomagnetic Ap index.

[0080] In an optional embodiment, solar wind parameters in the initial geomagnetic activity dataset are feature-filtered, and the nonlinear correlation coefficient between solar wind parameters and the geomagnetic Ap index is calculated to obtain a subset of key solar wind driving parameters, including:

[0081] Calculate the mutual information value and Pearson correlation coefficient between each solar wind parameter and the geomagnetic Ap index in the initial geomagnetic activity dataset;

[0082] The kernel density estimation method is used to perform a nonlinear mapping between mutual information value and Pearson correlation coefficient to obtain a comprehensive correlation evaluation index;

[0083] The solar wind parameters are sorted in descending order based on the comprehensive correlation evaluation index, and the parameters that rank at the top of the preset proportion are selected as key driving factors.

[0084] The selected key driving factors constitute a subset of key solar wind driving parameters.

[0085] It should be noted that for each solar wind parameter in the initial dataset, the system calculates its mutual information value and Pearson correlation coefficient with the geomagnetic Ap index; the mutual information value reflects the nonlinear relationship between the solar wind parameter and the geomagnetic Ap index, while the Pearson correlation coefficient measures the linear relationship between them;

[0086] Since the relationship between mutual information and Pearson correlation coefficient can be complex, the system uses kernel density estimation to perform a nonlinear mapping on these two indices. Kernel density estimation helps the system to understand the relationship between these two indices more precisely and generates a comprehensive correlation evaluation index through this process. The purpose of this step is to more comprehensively assess the impact of each solar wind parameter on the geomagnetic Ap index.

[0087] The solar wind parameters are sorted according to the comprehensive correlation evaluation index. The sorting order is based on the correlation strength between each parameter and the geomagnetic Ap index. During the sorting, the system selects those parameters with strong comprehensive correlation as key driving factors.

[0088] From the sorting results, the system selects the parameters that are at the top; these parameters usually occupy a preset proportion, indicating that they are the most critical in geomagnetic activity prediction; ultimately, these selected parameters will form the so-called "critical solar wind driving parameter subset" for subsequent geomagnetic Ap index prediction and modeling.

[0089] In an optional embodiment, with minimizing signal reconstruction error as the optimization objective, the penalty factor and decomposition level of the variational mode decomposition algorithm are globally optimized based on a subset of key solar wind driving parameters to obtain the optimal variational mode decomposition parameters, including:

[0090] The range of the penalty factor and the range of the number of decomposition levels for the variational mode decomposition algorithm are defined as the search space;

[0091] Using a subset of key solar wind driving parameters as input, the geomagnetic Ap exponent sequence is reconstructed using the variational mode decomposition algorithm, and the root mean square error between the reconstructed sequence and the original sequence is calculated.

[0092] The root mean square error is used as the fitness function value, and a genetic algorithm is used to perform an iterative search within the search space.

[0093] In each iteration, the combination of penalty factor and decomposition level is updated through selection, crossover, and mutation operations;

[0094] When the change in the fitness function value over multiple generations is less than the preset convergence threshold, the iteration stops, and the current optimal penalty factor and decomposition level are output as the optimal variational mode decomposition parameters.

[0095] It should be noted that two key parameters are set for the Variational Mode Decomposition (VMD) algorithm: the penalty factor and the number of decomposition levels. These parameters determine the "flexibility" of the VMD algorithm in processing data and the level of detail in the decomposition. The algorithm will search within the range of these parameters to find the most suitable combination.

[0096] Using a subset of key solar wind driving parameters as input, the variational mode decomposition algorithm is applied to reconstruct the geomagnetic Ap exponential sequence. This means that the VMD algorithm attempts to decompose the geomagnetic data into multiple eigenmode functions and reconstruct a new signal sequence accordingly.

[0097] There will be some differences between the reconstructed signal sequence and the original geomagnetic Ap index sequence; the system measures the reconstruction quality by calculating the root mean square error (RMSE) between the two; the smaller the error, the better the reconstruction effect.

[0098] To find the optimal penalty factor and decomposition level, the system uses a genetic algorithm for global optimization. The genetic algorithm simulates the process of natural selection, employing operations such as selection, crossover, and mutation to iterate within a set search space. After each iteration, the system evaluates the fitness of the current combination of penalty factor and decomposition level, that is, it measures the quality of the current combination based on the reconstruction error.

[0099] During each iteration, the genetic algorithm continuously updates the combination of penalty factor and decomposition layer number until it finds the optimal parameter combination. When the change in fitness function (reconstruction error) is less than a preset convergence threshold for multiple generations, it indicates that the algorithm has converged, meaning there is no room for further improvement, and the iteration stops.

[0100] The optimal penalty factor and decomposition level obtained in the current iteration are output as the optimal variational mode decomposition parameters; these parameters will be used in the decomposition and prediction process of the geomagnetic Ap exponential sequence, thereby improving the accuracy of the model.

[0101] In an optional embodiment, the geomagnetic Ap exponent historical sequence data is adaptively decomposed using optimal variational mode decomposition parameters to separate high-frequency noise components and intrinsic mode functions of different frequencies, resulting in a multi-scale geomagnetic signal sequence, including:

[0102] A constrained variational problem for constructing a variational mode decomposition model using optimal variational mode decomposition parameters;

[0103] The constrained variational problem is solved by using the alternating direction multiplier method, which decomposes the historical sequence data of the geomagnetic Ap index into several eigenmode functions and a residual component.

[0104] Based on the center frequency of each intrinsic mode function, high-frequency random fluctuation components are identified as noise and eliminated, while low-frequency and mid-frequency components with clear physical meaning are retained.

[0105] The retained intrinsic mode functions are arranged in descending order of frequency to form a multi-scale geomagnetic signal sequence.

[0106] It should be noted that, based on the previously obtained optimal VMD parameters, a variational mode decomposition model is constructed; this is an optimization problem with constraints, aiming to decompose the original geomagnetic Ap exponential sequence into several intrinsic mode functions (IMFs) and a residual component; the goal of this step is to accurately separate the different frequency components in the signal.

[0107] The constrained variational problem is solved using the Alternating Direction Multiplier Method (ADMM). ADMM is a commonly used optimization algorithm that can effectively decompose complex problems into simpler subproblems and solve them step by step. Through this method, the system can decompose historical data of the geomagnetic Ap index into multiple intrinsic mode functions (representing signal components of different frequencies) and a residual component (representing the part that cannot be captured by IMFs).

[0108] After obtaining multiple intrinsic mode functions, the frequency range of each mode will be different. The system analyzes the center frequency of each mode and identifies the higher frequency intrinsic modes (usually high-frequency random fluctuations) as noise components and removes them. In this way, the system only retains those low-frequency and mid-frequency signal components with clear physical meaning, which are usually related to geomagnetic activity.

[0109] The retained intrinsic mode functions are sorted from high to low according to their center frequencies to form a multi-scale geomagnetic signal sequence. This sorting method reflects the changes of geomagnetic signals at different time scales and can help researchers better understand how different frequency components of geomagnetic activity interact and their impact on the geomagnetic Ap index.

[0110] In an optional embodiment, a multi-scale geomagnetic signal sequence is input into a long short-term memory network model to learn the nonlinear evolution law in the short term, and output a short-term prediction sequence of the geomagnetic Ap index for the next 1 to 3 days, including:

[0111] Construct a long short-term memory network model structure that includes an input layer, a hidden layer, and an output layer, and set the number of neurons and activation function in the hidden layer;

[0112] Using multi-scale geomagnetic signal sequences as training samples, the backpropagation algorithm is used to update the weights and biases of the long short-term memory network model over time.

[0113] During model training, a dropout mechanism is introduced to prevent overfitting, and an early stopping strategy is used to determine the optimal training rounds.

[0114] By using a trained long short-term memory network model to perform forward inference on the current multi-scale geomagnetic signal sequence, a short-term prediction sequence of the geomagnetic Ap index for the next 1 to 3 days is obtained.

[0115] It should be noted that an LSTM network model is constructed, which includes an input layer, a hidden layer, and an output layer. The input layer receives data from multi-scale geomagnetic signal sequences, the hidden layer performs calculations through a series of neurons to learn the time dependence and nonlinear evolution of the data, and the output layer outputs the predicted geomagnetic Ap index value for the next 1 to 3 days based on the learned information.

[0116] The number of neurons in the hidden layer and the activation function need to be set; the number of neurons determines the complexity of the model, too many may lead to overfitting, and too few may lead to insufficient learning ability; the activation function helps the network capture non-linear features in the data;

[0117] Multiscale geomagnetic signal sequences were used as input data to train the LSTM model. During training, the backpropagation algorithm was used to update the model's weights and biases. The backpropagation algorithm is a common optimization method that calculates the prediction error and propagates the error back to each layer in the network to adjust the weights and biases, making the model's predictions more and more accurate.

[0118] To prevent overfitting during training (i.e., the model performs well on training data but poorly on new data), a dropout mechanism is introduced. The dropout mechanism improves generalization ability by randomly ignoring the activation values ​​of some neurons during training, forcing the model not to over-rely on certain neurons.

[0119] An early stopping strategy is employed; during training, the model's performance may begin to decline after some epochs (i.e., overfitting); the early stopping strategy monitors the performance on the validation set and stops training when the performance no longer improves, ensuring that the model is not overtrained.

[0120] After training, the trained LSTM model is used to perform forward inference on the current multi-scale geomagnetic signal sequence. Forward inference means using the model to generate output instead of updating parameters through backpropagation. In this step, the model will predict the geomagnetic Ap index sequence for the next 1 to 3 days based on historical data.

[0121] In an optional embodiment, a gated recurrent unit neural network model is constructed, and the hidden layer state of the model is initialized according to the 27-day recurrence cycle characteristics of geomagnetic activity to obtain a medium- to long-term prediction model, including:

[0122] Construct a gated recurrent unit neural network model and set up a memory unit inside the network specifically for storing periodic information;

[0123] Based on the 27-day recurrence period in the historical sequence data of the geomagnetic Ap index, the average period length and phase shift are calculated.

[0124] The initial hidden state of the gated recurrent unit neural network is assigned values ​​using the average period length and phase offset, so that the model has prior knowledge of the next 27-day cycle at the start of prediction.

[0125] The assigned gated recurrent unit neural network model is used as a medium- to long-term prediction model with a periodic memory mechanism.

[0126] It should be noted that a gated recurrent unit (GRU) neural network model is constructed. GRU is a variant of recurrent neural network (RNN) similar to LSTM, but with a simpler structure and higher computational efficiency. The key to GRU is its gating mechanism, which can automatically learn and remember long-term dependencies in time series data.

[0127] Within the GRU network, dedicated memory units are set up to store periodic information; these memory units can effectively capture the periodic characteristics of geomagnetic activity, especially the 27-day recurrence cycle, enabling the model to better process periodic data.

[0128] Based on historical data of the geomagnetic Ap index, its 27-day recurrence cycle was analyzed; this cycle reflects the regular fluctuations of geomagnetic activity within a specific time period.

[0129] The average period length and phase shift are calculated from historical data; the average period length represents the standard period of geomagnetic activity fluctuations, while the phase shift describes the deviation between the start time of each period and the average period; with these parameters, the model can more accurately understand the temporal patterns of geomagnetic activity.

[0130] The calculated average period length and phase offset are used as prior knowledge to assign values ​​to the initial hidden state of the GRU network. The purpose of this is to enable the model to "remember" the 27-day periodicity of geomagnetic activity at the start of the prediction. This assignment process provides the model with a memory of periodic data, enabling the model to make predictions based on this period from the beginning.

[0131] The assigned GRU neural network is used as a medium- to long-term prediction model with a periodic memory mechanism. This means that the model can not only learn short-term changes in geomagnetic activity, but also remember and use periodic features (i.e., a 27-day cycle) to make predictions for longer periods. Therefore, when making medium- to long-term predictions, the model will combine short-term trends and periodic fluctuations in historical data to provide more accurate predictions.

[0132] In an optional embodiment, a medium- to long-term prediction model is used to extract features from historical geomagnetic Ap index data to capture the periodic evolution trend caused by coronal hole high-velocity flow, and output a medium- to long-term prediction sequence of geomagnetic Ap index for the next few weeks to months, including:

[0133] Input historical data of geomagnetic Ap index into a medium- to long-term prediction model with a periodic memory mechanism according to the time step.

[0134] The retention and forgetting of historical information are controlled by using update gates and reset gates, with a focus on strengthening the memory weight of the 27-day cycle and its harmonic components;

[0135] The hidden layer state is mapped to the geomagnetic Ap index prediction for the next few weeks to months through a fully connected layer;

[0136] By continuously rolling the forecast of future time steps, a complete medium- and long-term forecast sequence of the geomagnetic Ap index is obtained by splicing the results.

[0137] It should be noted that the historical sequence data of the geomagnetic Ap index is input into a medium- to long-term prediction model with a periodic memory mechanism according to the time step; the "time step" here refers to the time interval between the data input to the model; for example, the model can take the geomagnetic Ap index data as input every day, every hour, etc., and train and predict based on the historical data.

[0138] In the recurrent units of the model, a gating mechanism is used to control the retention and forgetting of historical information. Specifically, the gating unit decides whether to retain or forget the previous state based on the current input data and historical information. Through the update gate (which determines whether to update the hidden state) and the reset gate (which determines whether to reset the memory state), the model can flexibly capture and remember important historical information while discarding parts that are no longer useful.

[0139] In this process, the key is to strengthen the memory weights of the 27-day cycle and harmonic components of the geomagnetic Ap index; because geomagnetic activity is periodic, especially the 27-day recurrence cycle and its harmonic components, these characteristics are crucial for accurately predicting future geomagnetic activity; therefore, the model will pay special attention to learning this periodic information.

[0140] After the hidden layer's state is calculated by the model, it is mapped to the geomagnetic Ap index prediction value for the next few weeks to months through a fully connected layer (i.e., a neural network layer). The fully connected layer converts the time series features learned by the hidden layer into specific prediction results, and finally outputs the future geomagnetic Ap index sequence.

[0141] Once the model has completed its predictions for the initial data, it will perform rolling predictions; this means that after each prediction of a future time period, the model will use the latest prediction as input data for the next time step and continue to make predictions; in this way, the model is able to generate predicted data for future time points step by step.

[0142] This rolling process will continue until the time span predicted by the model (such as several weeks or months) is fully generated, thus obtaining a complete long-term prediction sequence of the geomagnetic Ap index.

[0143] In an optional embodiment, based on the forecast lead time requirement, the short-term and medium-to-long-term forecast sequences of the geomagnetic Ap index are weighted and fused. An adaptive weight allocation strategy is used to coordinate the forecast results at different time scales to obtain the final intelligent forecast result of the geomagnetic Ap index, including:

[0144] Obtain the estimated time from the current forecast time for the solar wind shock to reach Earth, as well as the probability of a geomagnetic storm occurring;

[0145] Based on the estimated time and the probability of geomagnetic storm occurrence, the confidence weights of the short-term and medium-to-long-term prediction branches are dynamically calculated using fuzzy logic.

[0146] If the period is characterized by short-term disturbances, increase the weighting coefficient of the short-term forecast sequence; if the period is characterized by calm, increase the weighting coefficient of the medium- and long-term forecast sequences.

[0147] The weighted short-term and medium-to-long-term prediction sequences are linearly superimposed to obtain the final intelligent prediction result of the geomagnetic Ap index.

[0148] It should be noted that the model obtains both the current forecast time and the estimated time when the solar wind shock will reach Earth. This is because the arrival time of the solar wind shock affects geomagnetic activity, especially the occurrence of geomagnetic storms. At the same time, the model also needs to estimate the probability of geomagnetic storms. These two pieces of information can help determine how the prediction results at different time scales should be weighted.

[0149] Based on the estimated arrival time of the solar wind shock wave and the probability of geomagnetic storms, fuzzy logic is used to dynamically calculate the confidence weights of the short-term and medium-to-long-term prediction branches. The fuzzy logic system can handle uncertainty and fuzziness, and dynamically adjust the input information within a certain range, thereby calculating a confidence weight that adapts to the current situation.

[0150] These confidence weights represent the model's level of confidence in the prediction results at different time scales, i.e., under current conditions, whether short-term or medium- to long-term predictions contribute more to the final result.

[0151] If the current period is one of short-term disturbances (i.e., when geomagnetic activity is relatively intense), the model will increase the weight of short-term prediction sequences; because in this case, geomagnetic activity changes more drastically in the short term, and short-term prediction results are more important.

[0152] If the current period is a period of calm (i.e., geomagnetic activity is relatively stable), then the weight of medium- and long-term forecast sequences should be increased; at this time, the fluctuations in geomagnetic activity are smaller, and the long-term forecast results have a greater guiding effect on future trends.

[0153] Based on the calculated weight coefficients, the short-term prediction sequence and the medium- and long-term prediction sequence are weighted and fused. This means that, depending on the current level of disturbance activity or the quiet period, the model will assign different weights to the prediction results at different time scales, so that the final prediction results are more in line with the actual situation of the current geomagnetic activity.

[0154] This weighted fusion can be accomplished through linear superposition, that is, by weighting and summing the short-term and medium-to-long-term forecast results to obtain the final intelligent forecast result of the geomagnetic Ap index.

[0155] In an optional embodiment, the confidence weights of the short-term and medium-to-long-term prediction branches are dynamically calculated using fuzzy logic, including:

[0156] The fluctuation variance of the geomagnetic Ap index and the rate of change of solar wind speed are used as input variables for fuzzy logic;

[0157] Define a fuzzy set that reflects the uncertainty of prediction, and establish a fuzzy rule table between input variables and confidence weights;

[0158] Based on the current input variable values, the specific weight values ​​are calculated using the defuzzification method.

[0159] The calculated weight values ​​are normalized to ensure that the sum of the short-term forecast weight and the medium- and long-term forecast weight is the same.

[0160] It should be noted that, in order to dynamically calculate the confidence weight of the prediction results, some input variables are selected to reflect the state of geomagnetic activity. These variables include: the variance of fluctuation, which measures the degree of change in geomagnetic activity over a period of time and reflects the volatility and uncertainty of geomagnetic activity; and the change in the speed of the solar wind, which directly affects geomagnetic activity, and the rate of change of the solar wind speed, which can help determine whether there will be a violent geomagnetic disturbance.

[0161] Define some fuzzy sets to reflect the different states of input variables; for example, the variance of the geomagnetic Ap index can be divided into three states: "low volatility", "medium volatility" and "high volatility"; similarly, the rate of change of solar wind speed can also be divided into several states: "slow change", "moderate change" and "rapid change".

[0162] The fuzzy rule table associates different combinations of these input variables with corresponding confidence weights. Through these rules, the fuzzy logic system can dynamically adjust the confidence of short-term and medium-to-long-term predictions based on the current changes in geomagnetic fluctuations and solar wind speed. For example, when the geomagnetic Ap index fluctuates greatly and the solar wind speed changes rapidly, the system may increase the weight of short-term predictions because in this case, short-term predictions are more likely to reflect recent dramatic changes.

[0163] Once the values ​​of the input variables and the fuzzy rules are determined, the system uses a defuzzification method to calculate specific weight values. The defuzzification process transforms the fuzzy values ​​in the fuzzy set into clear values, thereby assigning a confidence weight to each prediction branch (short-term prediction and medium- to long-term prediction). For example, the weight of short-term prediction may be a value such as 0.7, indicating that the short-term prediction has high reliability in the current situation, while the weight of medium- to long-term prediction may be 0.3, indicating that its reliability is low.

[0164] The calculated weight values ​​need to be normalized to ensure that their sum is 1; because the weights of short-term and medium-to-long-term forecasts are complementary, normalization ensures their proportional relationship; in this way, the final weighted result can reasonably reflect the relative importance of the two forecast branches.

[0165] Example 2, please refer to Figure 2This invention provides a technical solution: a smart forecasting system for geomagnetic Ap index at multiple time scales, applicable to the aforementioned smart forecasting method for geomagnetic Ap index at multiple time scales, comprising:

[0166] Feature filtering module 1 is used to acquire historical sequence data of geomagnetic Ap index and corresponding real-time monitoring data of solar wind parameters, and to perform time-series alignment of the historical sequence data of geomagnetic Ap index to obtain an initial geomagnetic activity dataset; to perform feature filtering on solar wind parameters in the initial geomagnetic activity dataset, to calculate the nonlinear correlation coefficient between solar wind parameters and geomagnetic Ap index, and to obtain a subset of key solar wind driving parameters;

[0167] Mode decomposition module 2 is used to optimize the penalty factor and decomposition level of the variational mode decomposition algorithm based on a subset of key solar wind driving parameters, with the goal of minimizing signal reconstruction error, to obtain the optimal variational mode decomposition parameters.

[0168] Noise separation module 3 is used to adaptively decompose the geomagnetic Ap index historical sequence data using the optimal variational mode decomposition parameters, separate high-frequency noise components and intrinsic mode functions of different frequencies, and obtain multi-scale geomagnetic signal sequences.

[0169] The index prediction module 4 is used to input multi-scale geomagnetic signal sequences into a long short-term memory network model, learn the nonlinear evolution law in the short term, and output a short-term prediction sequence of the geomagnetic Ap index for the next 1 to 3 days; construct a gated recurrent unit neural network model, and initialize the hidden layer state of the model according to the 27-day recurrence cycle characteristics of geomagnetic activity to obtain a medium- and long-term prediction model.

[0170] Feature extraction module 5 is used to extract features from historical geomagnetic Ap index data using a medium- and long-term prediction model, capture the periodic evolution trend caused by coronal hole high-speed flow, and output the medium- and long-term prediction sequence of geomagnetic Ap index for the next few weeks to months.

[0171] The weighted fusion module 6 is used to perform weighted fusion of the short-term and medium-to-long-term prediction sequences of the geomagnetic Ap index according to the forecast timeliness requirements. It uses an adaptive weight allocation strategy to coordinate the prediction results at different time scales and obtain the final intelligent forecast result of the geomagnetic Ap index.

[0172] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A smart forecasting method for geomagnetic Ap index across multiple time scales, characterized in that, include: Historical sequence data of geomagnetic Ap index and corresponding real-time monitoring data of solar wind parameters were acquired, and the historical sequence data of geomagnetic Ap index were time-series aligned to obtain an initial geomagnetic activity dataset. Feature filtering was performed on the solar wind parameters in the initial geomagnetic activity dataset, and the nonlinear correlation coefficient between solar wind parameters and geomagnetic Ap index was calculated to obtain a subset of key solar wind driving parameters. With minimizing signal reconstruction error as the optimization objective, the penalty factor and decomposition level of the variational mode decomposition algorithm are globally optimized based on a subset of key solar wind driving parameters to obtain the optimal variational mode decomposition parameters; The geomagnetic Ap exponent historical sequence data is adaptively decomposed using the optimal variational mode decomposition parameters to separate high-frequency noise components and eigenmode functions of different frequencies, thereby obtaining a multi-scale geomagnetic signal sequence. Multi-scale geomagnetic signal sequences are input into a long short-term memory network model to learn the nonlinear evolution law in the short term and output a short-term prediction sequence of the geomagnetic Ap index for the next 1 to 3 days. A gated recurrent unit neural network model is constructed, and the hidden layer state of the model is initialized according to the 27-day recurrence cycle characteristics of geomagnetic activity to obtain a medium- and long-term prediction model. By using a medium- to long-term prediction model, features are extracted from historical geomagnetic Ap index data to capture the periodic evolution trend caused by high-speed flow in coronal holes, and a medium- to long-term prediction sequence of geomagnetic Ap index for the next few weeks to months is output. Based on the forecast timeliness requirements, the short-term and medium-to-long-term forecast sequences of the geomagnetic Ap index are weighted and fused. An adaptive weight allocation strategy is used to coordinate the forecast results at different time scales to obtain the final intelligent forecast result of the geomagnetic Ap index.

2. The intelligent forecasting method for geomagnetic Ap index oriented to multiple time scales according to claim 1, characterized in that, Feature filtering was performed on solar wind parameters in the initial geomagnetic activity dataset, and the nonlinear correlation coefficient between solar wind parameters and the geomagnetic Ap index was calculated to obtain a subset of key solar wind driving parameters, including: Calculate the mutual information value and Pearson correlation coefficient between each solar wind parameter and the geomagnetic Ap index in the initial geomagnetic activity dataset; The kernel density estimation method is used to perform a nonlinear mapping between mutual information value and Pearson correlation coefficient to obtain a comprehensive correlation evaluation index; The solar wind parameters are sorted in descending order based on the comprehensive correlation evaluation index, and the parameters that rank at the top of the preset proportion are selected as key driving factors. The selected key driving factors constitute a subset of key solar wind driving parameters.

3. The intelligent forecasting method for geomagnetic Ap index oriented to multiple time scales according to claim 2, characterized in that, With minimizing signal reconstruction error as the optimization objective, the penalty factor and decomposition level of the variational mode decomposition algorithm are globally optimized based on a subset of key solar wind driving parameters to obtain the optimal variational mode decomposition parameters, including: The range of the penalty factor and the range of the number of decomposition levels for the variational mode decomposition algorithm are defined as the search space; Using a subset of key solar wind driving parameters as input, the geomagnetic Ap exponent sequence is reconstructed using the variational mode decomposition algorithm, and the root mean square error between the reconstructed sequence and the original sequence is calculated. The root mean square error is used as the fitness function value, and a genetic algorithm is used to perform an iterative search within the search space. In each iteration, the combination of penalty factor and decomposition level is updated through selection, crossover, and mutation operations; When the change in the fitness function value over multiple generations is less than the preset convergence threshold, the iteration stops, and the current optimal penalty factor and decomposition level are output as the optimal variational mode decomposition parameters.

4. The intelligent forecasting method for geomagnetic Ap index oriented to multiple time scales according to claim 3, characterized in that, Adaptive decomposition of historical geomagnetic Ap exponent data is performed using optimal variational mode decomposition parameters to separate high-frequency noise components and eigenmode functions of different frequencies, resulting in a multi-scale geomagnetic signal sequence, including: A constrained variational problem for constructing a variational mode decomposition model using optimal variational mode decomposition parameters; The constrained variational problem is solved by using the alternating direction multiplier method, which decomposes the historical sequence data of the geomagnetic Ap index into several eigenmode functions and a residual component. Based on the center frequency of each intrinsic mode function, high-frequency random fluctuation components are identified as noise and eliminated, while low-frequency and mid-frequency components with clear physical meaning are retained. The retained intrinsic mode functions are arranged in descending order of frequency to form a multi-scale geomagnetic signal sequence.

5. The intelligent forecasting method for geomagnetic Ap index oriented to multiple time scales according to claim 4, characterized in that, Multi-scale geomagnetic signal sequences are input into a long short-term memory network model to learn the nonlinear evolution patterns in the short term, outputting a short-term prediction sequence of the geomagnetic Ap index for the next 1 to 3 days, including: Construct a long short-term memory network model structure that includes an input layer, a hidden layer, and an output layer, and set the number of neurons and activation function in the hidden layer; Using multi-scale geomagnetic signal sequences as training samples, the backpropagation algorithm is employed to update the weights and biases of the long short-term memory network model over time. During model training, a dropout mechanism is introduced to prevent overfitting, and an early stopping strategy is used to determine the optimal training rounds. By using a trained long short-term memory network model to perform forward inference on the current multi-scale geomagnetic signal sequence, a short-term prediction sequence of the geomagnetic Ap index for the next 1 to 3 days is obtained.

6. The intelligent forecasting method for geomagnetic Ap index oriented to multiple time scales according to claim 5, characterized in that, A gated recurrent unit neural network model is constructed, and the hidden layer states of the model are initialized based on the 27-day recurrence cycle characteristics of geomagnetic activity to obtain a medium- to long-term prediction model, including: Construct a gated recurrent unit neural network model and set up a memory unit inside the network specifically for storing periodic information; Based on the 27-day recurrence period in the historical sequence data of the geomagnetic Ap index, the average period length and phase shift are calculated. The initial hidden state of the gated recurrent unit neural network is assigned values ​​using the average period length and phase offset, so that the model has prior knowledge of the next 27-day cycle at the start of prediction. The assigned gated recurrent unit neural network model is used as a medium- to long-term prediction model with a periodic memory mechanism.

7. The intelligent forecasting method for geomagnetic Ap index oriented to multiple time scales according to claim 6, characterized in that, Feature extraction is performed on historical geomagnetic Ap index data using a medium- to long-term prediction model to capture the periodic evolution trend caused by high-velocity coronal hole flow, and to output medium- to long-term prediction sequences of geomagnetic Ap index for the next few weeks to months, including: Input historical data of geomagnetic Ap index into a medium- to long-term prediction model with a periodic memory mechanism according to the time step. The retention and forgetting of historical information are controlled by using update gates and reset gates, with a focus on strengthening the memory weight of the 27-day cycle and its harmonic components; The hidden layer state is mapped to the geomagnetic Ap index prediction for the next few weeks to months through a fully connected layer; By continuously rolling the forecast of future time steps, a complete medium- and long-term forecast sequence of the geomagnetic Ap index is obtained by splicing the results.

8. The intelligent forecasting method for geomagnetic Ap index oriented to multiple time scales according to claim 7, characterized in that, Based on the forecast lead time requirements, the short-term and medium-to-long-term forecast sequences of the geomagnetic Ap index are weighted and fused. An adaptive weight allocation strategy is used to coordinate the forecast results at different time scales, resulting in the final intelligent forecast result for the geomagnetic Ap index, including: Obtain the estimated time from the current forecast time for the solar wind shock to reach Earth, as well as the probability of a geomagnetic storm occurring; Based on the estimated time and the probability of geomagnetic storm occurrence, the confidence weights of the short-term and medium-to-long-term prediction branches are dynamically calculated using fuzzy logic. If the period is characterized by short-term disturbances, increase the weighting coefficient of the short-term forecast sequence; if the period is characterized by calm, increase the weighting coefficient of the medium- and long-term forecast sequences. The weighted short-term and medium-to-long-term prediction sequences are linearly superimposed to obtain the final intelligent prediction result of the geomagnetic Ap index.

9. The intelligent forecasting method for geomagnetic Ap index oriented to multiple time scales according to claim 8, characterized in that, The confidence weights of the short-term and medium-to-long-term prediction branches are dynamically calculated using fuzzy logic, including: The fluctuation variance of the geomagnetic Ap index and the rate of change of solar wind speed are used as input variables for fuzzy logic; Define a fuzzy set that reflects the uncertainty of prediction, and establish a fuzzy rule table between input variables and confidence weights; Based on the current input variable values, the specific weight values ​​are calculated using the defuzzification method. The calculated weight values ​​are normalized to ensure that the sum of the short-term forecast weight and the medium- and long-term forecast weight is the same.

10. A smart forecasting system for geomagnetic Ap index at multiple time scales, applicable to the smart forecasting method for geomagnetic Ap index at multiple time scales as described in any one of claims 1-9, characterized in that, include: The feature filtering module is used to acquire historical sequence data of geomagnetic Ap index and corresponding real-time monitoring data of solar wind parameters, and to perform time-series alignment of the historical sequence data of geomagnetic Ap index to obtain an initial geomagnetic activity dataset; to perform feature filtering on solar wind parameters in the initial geomagnetic activity dataset, to calculate the nonlinear correlation coefficient between solar wind parameters and geomagnetic Ap index, and to obtain a subset of key solar wind driving parameters; The mode decomposition module is used to optimize the signal reconstruction error by globally optimizing the penalty factor and the number of decomposition layers of the variational mode decomposition algorithm based on a subset of key solar wind driving parameters, so as to obtain the optimal variational mode decomposition parameters. The noise separation module is used to adaptively decompose the historical sequence data of the geomagnetic Ap index using the optimal variational mode decomposition parameters, separate high-frequency noise components and eigenmode functions of different frequencies, and obtain multi-scale geomagnetic signal sequences. The index prediction module is used to input multi-scale geomagnetic signal sequences into a long short-term memory network model, learn the nonlinear evolution law in the short term, and output a short-term prediction sequence of the geomagnetic Ap index for the next 1 to 3 days; construct a gated recurrent unit neural network model, and initialize the hidden layer state of the model according to the 27-day recurrence cycle characteristics of geomagnetic activity to obtain a medium- and long-term prediction model. The feature extraction module is used to extract features from historical geomagnetic Ap index data using a medium- to long-term prediction model, capture the periodic evolution trend caused by high-speed flow in coronal holes, and output medium- to long-term prediction sequences of geomagnetic Ap index for the next few weeks to months. The weighted fusion module is used to perform weighted fusion of short-term and medium-to-long-term prediction sequences of the geomagnetic Ap index according to the forecast timeliness requirements. It uses an adaptive weight allocation strategy to coordinate the prediction results at different time scales and obtain the final intelligent forecast result of the geomagnetic Ap index.

Images