Optical power prediction method and system based on multi-site spatial-temporal characteristics and dynamic optimization

Abstract

The invention discloses an optical power prediction method based on multi-site spatio-temporal characteristics and dynamic optimization. The method comprises the following steps: collecting data of a plurality of monitoring sites and carrying out characteristic selection by using a selection-deletion characteristic selection method; carrying out weighted fusion by adopting a multi-site data weighted fusion strategy; constructing an interval prediction model, inputting data subjected to feature selection and weighted fusion into an STG-Mamb-neural network model, and then performing LUBE interval prediction on output to obtain upper and lower boundaries of optical power; introducing an improved weighted average optimization algorithm, and optimizing the interval prediction model in combination with random disturbance and a boundary constraint strategy; and evaluating the optimized interval prediction model by adopting a cross validation method, and predicting the optical power of the multiple sites after the optimal interval prediction model is obtained. According to the invention, stable prediction performance can be maintained in a changeable environment, so that the reliability and stability of optical power prediction are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of optical power prediction, and in particular to optical power interval prediction in a photovoltaic power generation system. Background Art

[0002] Traditional optical power prediction methods primarily rely on meteorological data and historical power records, employing statistical models or simple machine learning algorithms. While these methods can provide optical power estimates to a certain extent, they often suffer from insufficient accuracy and poor generalization when faced with complex meteorological conditions and changing environmental factors. Furthermore, existing methods mostly focus on point predictions and lack the ability to quantify prediction uncertainty, making them incapable of providing reliable decision-making for grid scheduling and energy management.

[0003] Most existing studies focus solely on data from a single site, ignoring the spatial correlation of optical power across different geographic locations and environmental conditions. Furthermore, training deep learning models typically requires extensive data and computing resources, with high data quality requirements, which can be limiting in practical applications.

[0004] To overcome the limitations of point prediction, interval prediction methods have emerged. Existing interval prediction methods are mostly based on simple statistical models or single machine learning algorithms, which struggle to effectively handle complex spatiotemporal data. Regarding multi-site data fusion, while some studies have attempted to integrate data from different sites through weighting or splicing, these methods often lack comprehensive consideration of data quality, geographic location, and environmental factors, resulting in the fused data failing to fully realize its value. Summary of the Invention

[0005] Purpose of the invention: The present invention aims to provide a high-precision and high-reliability optical power interval prediction method, which overcomes the problems of insufficient prediction accuracy, weak generalization ability and lack of uncertainty quantification in the existing technology by integrating multi-site spatiotemporal data, deep learning models and dynamic optimization algorithms.

[0006] Technical solution: An optical power prediction method based on multi-site spatiotemporal characteristics and dynamic optimization includes the following steps:

[0007] Step 1: Collect optical power data, cloud image data, and meteorological data from multiple monitoring stations, and use the select-drop feature selection method to select the features most relevant to optical power prediction;

[0008] Step 2: Using a multi-site data weighted fusion strategy, optical power data, cloud image data, and meteorological data are weighted and fused according to the error, geographical location, and environmental factors of each monitoring site;

[0009] Step 3: Construct an interval prediction model that combines the STG-Mamba neural network model and the LUBE method. Input the feature-selected and weighted fusion data into the STG-Mamba neural network model and then predict the LUBE interval of the output to obtain the upper and lower bounds of the optical power.

[0010] Step 4: Introduce the improved weighted average optimization algorithm, combine random perturbation and boundary constraint strategies, and optimize the interval prediction model;

[0011] Step 5: Use the cross-validation method to evaluate the optimized interval prediction model, and predict the optical power of multiple sites after obtaining the optimal interval prediction model.

[0012] Furthermore, the select-delete feature selection method comprises the following steps:

[0013] (1) The importance of features is evaluated using the minimum redundancy maximum relevance (mRMR) criterion; this method maximizes the correlation between features and target variables (correlation term) while minimizing the redundancy between features (redundancy term), thereby screening out a representative feature subset. For candidate feature f i , its importance score can be expressed as:

[0014]

[0015] Among them, y is the target variable, S is the currently selected feature subset, I(f i ; y) represents feature f i The mutual information with the target variable y measures its predictive ability; I(f i ;f j ) represents the feature f i With the selected feature f j The mutual information of X and Y is used to measure their redundancy. The calculation formula of mutual information I(X; Y) is:

[0016]

[0017] Here, p(x, y) is the joint probability distribution, and p(x) and p(y) are the marginal probability distributions. This scoring mechanism retains the top-scoring features, constructing a feature subset with high information content and low redundancy, thereby improving the generalization and predictive performance of subsequent models.

[0018] (2) The exploration strategy in the feature selection process adopts the position update rule of the weighted average optimization algorithm;

[0019] The WAA exploration strategy guides the search direction by weightedly fusing the current positions of multiple individuals, achieving efficient exploration of the feature subspace. This strategy leverages information interaction between individuals to update the search, offering excellent global search capabilities and convergence stability, making it suitable for selection tasks involving large or high-dimensional feature spaces.

[0020] (3) Selection-deletion operation: By adjusting the probability of deleting features (DR) and the probability of selecting features (SR), a balance is found between global search and local search. In the initial stage, the probability of deleting features is higher, which helps to remove redundant and unimportant features and conduct global search. In the later stage, the probability of selecting features increases, further refining the search and focusing on the local optimal solution.

[0021] The rate of selection and deletion is controlled by a formula that ensures that the entire process can both widely explore the feature space and locally optimize key features in the final stage. The formula is as follows:

[0022] The probability of deleting a feature DR and the probability of selecting a feature SR are dynamically adjusted in each iteration:

[0023]

[0024] SR=1-DR

[0025] Among them, DR max and DR min is the maximum and minimum probability of deleting features, MaxFEs is the maximum number of evaluations, and FEs is the current number of evaluations. By adjusting the deletion and selection rates, the algorithm finds a balance between global search and local optimization.

[0026] Furthermore, the multi-site data weighted fusion strategy is specifically: considering data quality, geographical location and environmental factors to fuse the optical power data of multiple sites. The formula is as follows:

[0027]

[0028] Among them, D 融合 Denotes the fused dataset, D i represents the dataset of the i-th site, w i represents the weight of the i-th site data, and n represents the total number of sites;

[0029]

[0030] in: They represent the weights based on data quality, geographical location, and environmental factors, respectively. α, β, and γ are hyperparameters that can be adjusted according to actual needs.

[0031] The data quality-based weight takes into account the historical errors of each site's data and assigns higher weights to sites with smaller errors. The specific calculation formula is:

[0032]

[0033] Among them, MSE i represents the mean square error of site i. Data with smaller mean square error are more reliable, so their corresponding weight is higher.

[0034] The weights based on geographic location, if there is a strong spatial correlation between sites, the monitoring data located close to the target site should be given a higher weight; the Gaussian kernel function is used to calculate the spatial correlation:

[0035]

[0036] Among them, x i and x j denote the geographical coordinates of site i and target site j, σ spatial It is a parameter that controls the degree of influence of spatial distance on correlation, called spatial bandwidth;

[0037] The calculated K(x i ,x j ) value is directly used as the weight of site i to the target site j data;

[0038] To avoid excessive or insufficient weights, the weights based on geographic location are normalized so that the sum of the weights of all sites is 1. The normalized weight formula is:

[0039]

[0040] In this way, the weight reflects the contribution of each site in the prediction process based on its spatial location. Sites with closer distances will be given higher weights, thereby improving the prediction accuracy.

[0041] The weight based on environmental factors considers the impact of environmental conditions on optical power, including temperature, humidity, and wind speed. The weighting is based on the environmental similarity between the site and the target site. The specific calculation formula is as follows:

[0042]

[0043] Among them, T j ,H j ,V j is the temperature, humidity, and wind speed of the target site, T i ,H i ,V i is the temperature, humidity, and wind speed at the i-th station.

[0044] Furthermore, a multi-site data weighted fusion strategy is used to pre-process the data after fusion, including data cleaning and standardization. Data cleaning removes outliers and erroneous records; standardization is performed using the following formula:

[0045]

[0046] Where X is the original data, μ is the mean, σ is the standard deviation, and Z is the standardized data.

[0047] Furthermore, the interval prediction model includes:

[0048] (1) Inputting the cloud image, time series optical power data and meteorological data after feature selection and weighted fusion into the STG-Mamba model;

[0049] (2) The predicted value of the optical power output by the STG-Mamba model and its uncertainty interval are used as the input of the LUBE method;

[0050] (3) Use the LUBE method to perform interval estimation on the optical power prediction results;

[0051] (4) The indicators of interval coverage loss PICP and interval width loss PIWL are introduced into the LUBE method; and the interval prediction loss function CWCM is proposed, which is as follows:

[0052]

[0053] Among them, λ1 and λ2 are weighting coefficients, which are used to adjust the relative importance of PICP and PIWL respectively; η1 and η2 are attenuation factors, which are used to control the influence of PICP and PIWL on the objective function.

[0054] Furthermore, the step 4 introduces an improved weighted average optimization algorithm to optimize the key parameters of the interval prediction model so that the interval prediction loss function CWCM reaches the optimal value, including the following steps:

[0055] (1) Initialize parameter population

[0056] In the weighted average optimization algorithm, first initialize N individuals in the population, that is, the parameters of N groups of interval prediction models:

[0057]

[0058]

[0059] Where n represents the nth individual, i.e. the nth set of hyperparameters, k represents the kth hyperparameter, and d represents the total number of optimized hyperparameters; and represents the boundary of the kth hyperparameter, rand is a uniform random number between [0,1], ensuring that the initial population is evenly distributed in the search space;

[0060] (2) Calculate the objective function

[0061] For each set of hyperparameters θ n , train the interval prediction model and calculate the loss function CWCM:

[0062]

[0063] Among them, λ1 and λ2 are weighting coefficients used to adjust the relative importance of PICP and PIWL respectively; η1 and η2 are attenuation factors used to control the influence of PICP and PIWL on the objective function;

[0064] Compute the weighted center of the current hyperparameter population:

[0065]

[0066] Among them, Fitness(θ i ) represents the fitness of the nth candidate solution, θ center Represents the weighted average position, N sel is the number of candidate solutions selected;

[0067] (3) Introducing random perturbations and boundary constraints:

[0068] 3.1 Local Search

[0069]

[0070] Among them, w1 and w2 are adaptive weights, randn obeys the standard normal distribution, ∈ is the disturbance factor, ∈0 is the initial disturbance amplitude, T max is the maximum number of iterations; t is the number of iterations;

[0071] 3.2 Global Search

[0072] In order to enhance the global search capability, a disturbance factor is introduced to perform global jumps:

[0073]

[0074] Among them, γ0 is the initial perturbation, γ controls the perturbation amplitude, and gradually converges with iteration;

[0075] 3.3 Boundary Constraints:

[0076] If the new hyperparameters calculated If the value exceeds the set range, it will be corrected to the nearest boundary value:

[0077]

[0078] In addition, the reflection boundary mechanism is combined to further optimize:

[0079]

[0080] When the hyperparameter exceeds the boundary, it is reflected to the legal area;

[0081] (4) The improved weighted average optimization algorithm terminates the optimization when any of the following conditions is met: when the difference between the loss function values of two consecutive iterations is less than the preset convergence threshold ∈ loss That is |CWCM(θ (t+1) 0-CWCM(θ (t) )|<∈ loss , indicating that the loss function has converged; or when the number of iterations reaches the preset maximum value T max , the algorithm stops optimizing further.

[0082] Through an improved weighted average optimization algorithm, the model can adjust key parameters based on the dynamic changes in prediction error, thereby optimizing the LUBE prediction interval. Ultimately, these optimization strategies work together to enable the model to more accurately characterize uncertainty in optical power prediction and improve the reliability of the prediction interval.

[0083] Furthermore, step five includes:

[0084] The K-fold cross-validation method is used to evaluate the optimized interval prediction model. First, the complete dataset is divided into K subsets of equal proportions. Then, in each round of training, K-1 subsets are selected as training sets, and the remaining subset is selected as the test set. The model is trained and the prediction error is calculated. The test set is rotated in sequence, and this is repeated K times so that all data participate in training and validation. Finally, the mean of all K rounds of test errors is used as the final evaluation indicator to obtain the optimal interval prediction model for multi-site optical power prediction.

[0085] The optical power prediction system based on multi-site spatiotemporal characteristics and dynamic optimization includes a data acquisition module, a data screening module, a data fusion module, an interval prediction model, and an optical power prediction module;

[0086] The data acquisition module collects optical power data, cloud image data and meteorological data of multiple monitoring sites;

[0087] The data screening module uses a selection-deletion feature selection method to perform feature selection to screen out the features most relevant to optical power prediction;

[0088] The data fusion module adopts a multi-site data weighted fusion strategy to perform weighted fusion of optical power data, cloud image data and meteorological data according to the error, geographical location and environmental factors of each monitoring site;

[0089] The interval prediction model combines the STG-Mamba-neural network model and the LUBE method; after inputting the feature-selected and weighted fusion data into the STG-Mamba-neural network model, the output is predicted by the LUBE interval to obtain the upper and lower boundaries of the optical power; an improved weighted average optimization algorithm is introduced, combined with random perturbation and boundary constraint strategies, to optimize the interval prediction model;

[0090] The optical power prediction module uses a cross-validation method to evaluate the optimized interval prediction model, and predicts the optical power of multiple sites after obtaining the optimal interval prediction model.

[0091] Beneficial Effects: (1) By fusing multi-site spatiotemporal data, the present invention effectively integrates optical power data from different geographical locations and environmental conditions. This data fusion strategy significantly enhances the model's adaptability to complex scenarios, enabling it to maintain stable prediction performance in a changing environment, thereby improving the reliability and stability of optical power prediction.

[0092] (2) By using a select-and-drop feature selection method, the present invention can accurately screen out the features most relevant to optical power prediction and remove redundant information. This process not only significantly improves data quality but also significantly enhances computational efficiency, providing high-quality input data for model training and ensuring that the model maintains high-precision prediction capabilities while operating efficiently.

[0093] (3) The interval prediction model combines the STG-Mamba neural network model and the LUBE interval prediction method to effectively capture the dynamic changes and spatial correlations of optical power. This combination not only enhances the model's sensitivity to optical power changes, but also significantly improves the prediction accuracy, enabling it to more accurately reflect the actual trend of optical power changes. The LUBE interval prediction further quantifies the uncertainty of the prediction, providing a more reliable decision-making basis for grid scheduling and energy management. The interval prediction not only provides an estimate of the optical power, but also provides a more comprehensive reference for practical applications by predicting the upper and lower boundaries, helping decision makers better evaluate and deal with the uncertainty in optical power prediction.

[0094] (4) The improved weighted average optimization algorithm (IWAA) was introduced. This method optimizes the model training process, accelerates convergence, and ensures the model's efficiency and stability. By dynamically adjusting the weights of the prediction results and combining random perturbations and boundary constraint strategies, IWAA further enhances the model's generalization and stability, enabling it to maintain good prediction performance across different datasets. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] Figure 1 This is a flow chart of the optical power prediction method based on multi-site spatiotemporal characteristics and dynamic optimization of the present invention;

[0096] Figure 2 Flowchart of the improved weighted average optimization algorithm (IWAA) of the present invention. DETAILED DESCRIPTION

[0097] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0098] like Figure 1 As shown, the optical power prediction method based on multi-site spatiotemporal characteristics and dynamic optimization includes the following steps:

[0099] Step 1: Collect optical power data, cloud image data, and meteorological data from multiple monitoring stations, and use the select-drop feature selection method to select the features most relevant to optical power prediction;

[0100] Step 2: Using a multi-site data weighted fusion strategy, optical power data, cloud image data, and meteorological data are weighted and fused according to the error, geographical location, and environmental factors of each monitoring site;

[0101] Step 3: Construct an interval prediction model that combines the STG-Mamba neural network model and the LUBE method. Input the feature-selected and weighted fusion data into the STG-Mamba neural network model and then predict the LUBE interval of the output to obtain the upper and lower bounds of the optical power.

[0102] Step 4: Introduce the improved weighted average optimization algorithm, combine random perturbation and boundary constraint strategies, and optimize the interval prediction model;

[0103] Step 5: Use the cross-validation method to evaluate the optimized interval prediction model, and predict the optical power of multiple sites after obtaining the optimal interval prediction model.

[0104] Furthermore, the implementation process of step 1 is as follows: First, environmental monitoring data from multiple sensors is collected, including but not limited to the following:

[0105] Optical power data comes from photovoltaic power stations or optical radiation measurement equipment, which records the optical power output at different time points; cloud image data is obtained through satellite remote sensing or ground-based camera equipment, and contains information such as cloud amount, cloud thickness, cloud type, and cloud movement speed. These factors directly affect the stability of solar radiation; meteorological data includes temperature, humidity, wind speed, atmospheric pressure, etc., and is provided by meteorological stations or online monitoring equipment to help comprehensively analyze the changing patterns of optical power.

[0106] Cloud image data is matched with meteorological data using timestamps and geographic coordinates to ensure temporal and spatial alignment with optical power data. To improve the model's generalization capabilities, the raw data may contain noise, redundant features, or outliers, requiring preprocessing steps such as data cleaning, noise removal, outlier detection, and missing data filling. Cleaning removes erroneous records and missing values. Noise removal uses smoothing methods, outlier detection processes abnormal data, and missing data filling restores missing data using interpolation or prediction methods. After preprocessing, feature selection is used to filter out features relevant to optical power prediction, remove redundant information, improve data quality and computational efficiency, and provide more accurate input data for model training.

[0107] Furthermore, the select-delete feature selection method comprises the following steps:

[0108] (1) The importance of features is evaluated using the minimum redundancy maximum relevance (mRMR) criterion; this method maximizes the correlation between features and target variables (correlation term) while minimizing the redundancy between features (redundancy term), thereby screening out a representative feature subset. For candidate feature f i , its importance score can be expressed as:

[0109]

[0110] Among them, y is the target variable, S is the currently selected feature subset, I(f i ; y) represents feature f i The mutual information with the target variable y measures its predictive ability; I(f i ;f j ) represents the feature f i With the selected feature f j The mutual information of X and Y is used to measure their redundancy. The calculation formula of mutual information I(X; Y) is:

[0111]

[0112] Here, p(x, y) is the joint probability distribution, and p(x) and p(y) are the marginal probability distributions. This scoring mechanism retains the top-scoring features, constructing a feature subset with high information content and low redundancy, thereby improving the generalization and predictive performance of subsequent models.

[0113] (2) The exploration strategy in the feature selection process adopts the position update rule of the weighted average optimization algorithm;

[0114] The WAA exploration strategy guides the search direction by weightedly fusing the current positions of multiple individuals, achieving efficient exploration of the feature subspace. This strategy leverages information interaction between individuals to update the search, offering excellent global search capabilities and convergence stability, making it suitable for selection tasks involving large or high-dimensional feature spaces.

[0115] (3) Selection-deletion operation: By adjusting the probability of deleting features (DR) and the probability of selecting features (SR), a balance is found between global search and local search. In the initial stage, the probability of deleting features is higher, which helps to remove redundant and unimportant features and conduct global search. In the later stage, the probability of selecting features increases, further refining the search and focusing on the local optimal solution.

[0116] The rate of selection and deletion is controlled by a formula that ensures that the entire process can both widely explore the feature space and locally optimize key features in the final stage. The formula is as follows:

[0117] The probability of deleting a feature DR and the probability of selecting a feature SR are dynamically adjusted in each iteration:

[0118]

[0119] SR=1-DR

[0120] Among them, DR max and DR min is the maximum and minimum probability of deleting features, MaxFEs is the maximum number of evaluations, and FEs is the current number of evaluations. By adjusting the deletion and selection rates, the algorithm finds a balance between global search and local optimization.

[0121] Because optical power is affected by multiple factors, including geography, weather variations, and equipment differences, data from different sites often differ. Therefore, to ensure effective collaboration between site data, a weighted fusion approach was adopted. This approach assigns different weights to each site based on its reliability and contribution, allowing it to fully leverage its role in the final forecast and ultimately improve overall prediction accuracy.

[0122] The multi-site data weighted fusion strategy is specifically: considering data quality, geographical location and environmental factors to fuse the optical power data of multiple sites. The formula is as follows:

[0123]

[0124] Among them, D 融合 Denotes the fused dataset, D i represents the dataset of the i-th site, w i represents the weight of the i-th site data, and n represents the total number of sites;

[0125]

[0126] in: They represent the weights based on data quality, geographical location, and environmental factors, respectively. α, β, and γ are hyperparameters that can be adjusted according to actual needs.

[0127] The data quality-based weight takes into account the historical error of each site's data (such as mean square error (MSE)) and assigns higher weights to sites with smaller errors. The specific calculation formula is:

[0128]

[0129] Among them, MSE i represents the mean square error of site i. Data with smaller mean square error are more reliable, so their corresponding weight is higher.

[0130] The weights based on geographic location, if there is a strong spatial correlation between sites, the monitoring data located close to the target site should be given a higher weight; the Gaussian kernel function is used to calculate the spatial correlation:

[0131]

[0132] Among them, x i and x j denote the geographical coordinates of site i and target site j, σ spatial It is a parameter that controls the degree of influence of spatial distance on correlation, called spatial bandwidth;

[0133] The calculated K(x i ,x j ) value is directly used as the weight of site i to the target site j data;

[0134] To avoid excessive or insufficient weights, the weights based on geographic location are normalized so that the sum of the weights of all sites is 1. The normalized weight formula is:

[0135]

[0136] In this way, the weight reflects the contribution of each site in the prediction process based on its spatial location. Sites with closer distances will be given higher weights, thereby improving the prediction accuracy.

[0137] The weight based on environmental factors considers the impact of environmental conditions on optical power, including temperature, humidity, and wind speed. The weighting is based on the environmental similarity between the site and the target site. The specific calculation formula is as follows:

[0138]

[0139] Among them, T j ,H j ,V j is the temperature, humidity, and wind speed of the target site, T i ,H i ,V i is the temperature, humidity, and wind speed at the i-th station.

[0140] Furthermore, a multi-site data weighted fusion strategy is used to pre-process the data after fusion, including data cleaning and standardization. Data cleaning removes outliers and erroneous records; standardization is performed using the following formula:

[0141]

[0142] Where X is the original data, μ is the mean, σ is the standard deviation, and Z is the standardized data.

[0143] Furthermore, the interval prediction model includes:

[0144] (1) Inputting the cloud image, time series optical power data and meteorological data after feature selection and weighted fusion into the STG-Mamba model;

[0145] (2) The predicted value of the optical power output by the STG-Mamba model and its uncertainty interval are used as the input of the LUBE method;

[0146] (3) Use the LUBE method to perform interval estimation on the optical power prediction results;

[0147] (4) The indicators of interval coverage loss PICP and interval width loss PIWL are introduced into the LUBE method; and the interval prediction loss function CWCM is proposed, which is as follows:

[0148]

[0149] Among them, λ1 and λ2 are weighting coefficients, which are used to adjust the relative importance of PICP and PIWL respectively; η1 and η2 are attenuation factors, which are used to control the influence of PICP and PIWL on the objective function.

[0150] STG-Mamba adopts a stacked encoder-decoder architecture, the core of which consists of a graph-selective state-space block (GS3B). Each GS3B consists of a spatiotemporal selective state-space module (ST-S3M) and a Kalman filter graph neural network (KFGN).

[0151] This architecture utilizes an encoder-decoder structure to process the input and output of time series data, making it suitable for modeling dynamic time series data. ST-S3M uses the Graph Selective Scan (GSSSM) mechanism to selectively extract features from cloud maps, optical power time series data, and meteorological data after RFGSDFS feature extraction, removing redundant information and optimizing spatiotemporal feature modeling. KFGN, combined with Kalman filtering theory, dynamically adjusts node relationships in the graph structure, facilitating modeling of data features in complex environments and improving model robustness.

[0152] The model inputs are cloud images that have undergone feature selection and weighted fusion, time-series optical power data, and meteorological data. The output is the predicted optical power value and its uncertainty interval. During model operation, the SSSM simulates and predicts the dynamic changes in system states, while the KFGN combines with the Kalman filter method to optimize the parameter estimation of the graph neural network, improving the accuracy of optical power prediction.

[0153] First, the cloud imagery, time-series optical power data, and meteorological data, which have undergone feature selection and weighted fusion, are fed into the STG-Mamba model. This model combines the Selective State Space Model (SSSM) with the Kalman Filter Graph Neural Network (KFGN) to effectively capture the dynamic changes and spatial correlations between different sites.

[0154] In the operation of the model, SSSM is mainly used to simulate and predict the dynamic changes of the system state, while KFGN is combined with the Kalman filtering method to optimize the parameter estimation of the graph neural network, thereby improving the accuracy of the model's optical power prediction.

[0155] SSSM adopts the selective state space modeling method and models the key spatiotemporal features through the Graph Selective Scan (GSSSM) mechanism to reduce redundant calculations and improve prediction efficiency. Assume that the hidden state of the system is h t , the input data (cloud map and time series data after feature extraction) is u t , then the state update equation is:

[0156] h t+1 =Ah t +Bu t +ε t

[0157] The observation equation describes the relationship between the observed value (optical power prediction value) and the state variable of the system:

[0158]

[0159] Among them: A, B, C are the state transition matrix, input mapping matrix and observation matrix, which are automatically learned by STG-Mamba through deep learning.

[0160] ε t ~N(0,Σ) is the process noise, the instability of the reaction system, δ t ~N(0,Γ) is the measurement noise, which represents the observation error.

[0161] On this basis, KFGN uses the Kalman filter mechanism to update the weights between nodes in the GNN structure, making the model more robust. Specifically, the Kalman filter performs noise suppression and state estimation on time series data through a two-step process: prediction and update.

[0162] Prediction steps:

[0163]

[0164] P t|t-1 =AP t-1|t-1 A T +Q

[0165] Update steps:

[0166] K t =P t|t-1 C T (CP t|t-1 C T +R) -1

[0167]

[0168] P t|t =(IK t C)P t|t-1

[0169] in, is the predicted value based on the previous state and input, P t|t-1 is the predicted covariance matrix, Q is the process noise covariance, reflecting the uncertainty of the system state change. t represents the Kalman gain, which controls the weight of state update. R is the measurement noise covariance.

[0170] Through the above process, KFGN adaptively adjusts the spatiotemporal relationship of optical power prediction by combining Kalman filtering and GNN, eliminates noise interference, and improves the stability and accuracy of prediction.

[0171] Spatial correlation modeling

[0172] Assume that the stations form an undirected graph G = (V, E), where the node set V represents the stations and the edge set E represents the optical power correlation between the stations. KFGN introduces Kalman filtering into the GNN structure, which can model and suppress the effects of noise, making the prediction results more robust.

[0173] Ultimately, the output of the STG-Mamba model (the predicted optical power value and its uncertainty interval) serves as input to the LUBE method, which further calculates and optimizes the optical power prediction interval. Specifically, STG-Mamba not only provides a point prediction value for optical power, but also provides the uncertainty interval for its prediction. Using these upper and lower bounds, the LUBE method can conduct decision analysis based on this information, helping to more accurately assess and address the uncertainty in optical power prediction, thereby providing more comprehensive and reliable decision support for practical applications.

[0174] Furthermore, to effectively quantify the uncertainty of optical power prediction and ensure that the prediction interval covers the true value, the present invention adopts the LowerUpperBoundEstimation (LUBE) method to perform interval estimation on the optical power prediction results. Through the LUBE method, the optical power prediction point value and related uncertainty information can be extracted from the STG-Mamba neural network model to dynamically estimate and optimize the prediction interval. The details are as follows:

[0175] Using the LUBE interval prediction method, the upper and lower bounds of the prediction interval are calculated using the following formula:

[0176]

[0177] in, represents the predicted optical power value at time t, σ(t) is the uncertainty (standard deviation) of the prediction, and α represents the confidence coefficient. When the confidence level is set to 95%, α = 1.96. The uncertainty σ(t) is determined by the process noise and measurement noise of the SSSM. Assuming that the process noise ∈ t and measurement noise δ t They are not correlated with each other, and the calculation formula is as follows:

[0178]

[0179] Where, process noise ∈ t and measurement noise δ t The Kalman filtering method is used to dynamically update the uncertainty estimation to ensure the accuracy.

[0180] Furthermore, to effectively quantify the uncertainty of optical power prediction and ensure that the prediction interval covers the true value, the present invention adopts the LowerUpperBoundEstimation (LUBE) method to perform interval estimation on the optical power prediction results. Through the LUBE method, the optical power prediction point value and related uncertainty information can be extracted from the STG-Mamba neural network model to dynamically estimate and optimize the prediction interval. The details are as follows:

[0181] Using the LUBE interval prediction method, the upper and lower bounds of the prediction interval are calculated using the following formula:

[0182] U(t)=y train (t)+(1+10%)

[0183] L(t)=y train (t)+(1-10%)

[0184] To comprehensively evaluate the quality and performance of prediction intervals, this paper introduces two key metrics: probability of interval coverage loss (PICP) and probability of interval width loss (PIWL). These two metrics measure the accuracy and compactness of the prediction interval from different perspectives. PICP assesses whether the actual value falls within the predicted interval, reflecting the reliability of the prediction interval; while PIWL focuses on the width of the prediction interval, reflecting the compactness of the prediction interval and further measuring the accuracy of the prediction result.

[0185] (1) Interval coverage ratio (PICP):

[0186]

[0187] in, is the indicator function, when the actual value y t When it falls within the prediction interval, otherwise

[0188] (2) Interval Width (PIWL):

[0189]

[0190] Where R is the range of optical power data. The smaller PINAW is, the tighter the prediction interval is and the better the prediction effect is.

[0191] In order to balance the two indicators of interval coverage (PICP) and interval width (PIWL), and ensure that the optimization target can comprehensively consider the coverage and width while avoiding excessive penalties, this paper proposes a comprehensive optimization loss function (CWCM) for interval coverage and interval width. This loss function flexibly adjusts the relative influence of PICP and PIWL in the optimization process by introducing an attenuation factor and a weighting coefficient, avoiding the excessive influence of a certain indicator on the final result. The formula of CWCM is as follows:

[0192]

[0193] Where λ1 and λ2 are weighting coefficients used to adjust the relative importance of PICP and PIWL, respectively. η1 and η2 are attenuation factors used to control the influence of PICP and PIWL on the objective function.

[0194] Further, such as Figure 2 As shown, the improved weighted average optimization algorithm optimizes the key parameters of the interval prediction model to optimize the interval prediction loss function CWCM, including the following steps:

[0195] (1) Initialize parameter population

[0196] In the weighted average optimization algorithm, first initialize N individuals in the population, that is, the parameters of N groups of interval prediction models:

[0197]

[0198] Where n represents the nth individual, i.e. the nth set of hyperparameters, k represents the kth hyperparameter, and d represents the total number of optimized hyperparameters; and represents the boundary of the kth hyperparameter, rand is a uniform random number between [0,1], ensuring that the initial population is evenly distributed in the search space;

[0199] (2) Calculate the objective function

[0200] For each set of hyperparameters θ n , train the interval prediction model and calculate the loss function CWCM:

[0201]

[0202] Among them, λ1 and λ2 are weighting coefficients used to adjust the relative importance of PICP and PIWL respectively; η1 and η2 are attenuation factors used to control the influence of PICP and PIWL on the objective function;

[0203] Compute the weighted center of the current hyperparameter population:

[0204]

[0205] Among them, Fitness(θ i ) represents the fitness of the nth candidate solution, θ center Represents the weighted average position, N sel is the number of candidate solutions selected;

[0206] (3) Introducing random perturbations and boundary constraints:

[0207] 3.1 Local Search

[0208]

[0209]

[0210] Among them, w1 and w2 are adaptive weights, randn obeys the standard normal distribution, ∈ is the disturbance factor, ∈0 is the initial disturbance amplitude, T max is the maximum number of iterations; t is the number of iterations;

[0211] 3.2 Global Search

[0212] In order to enhance the global search capability, a disturbance factor is introduced to perform global jumps:

[0213]

[0214] Among them, γ0 is the initial perturbation, γ controls the perturbation amplitude, and gradually converges with iteration;

[0215] 3.3 Boundary Constraints:

[0216] If the new hyperparameters calculated If the value exceeds the set range, it will be corrected to the nearest boundary value:

[0217]

[0218] In addition, the reflection boundary mechanism is combined to further optimize:

[0219]

[0220] When the hyperparameter exceeds the boundary, it is reflected to the legal area;

[0221] (4) The improved weighted average optimization algorithm terminates the optimization when any of the following conditions is met: when the difference between the loss function values of two consecutive iterations is less than the preset convergence threshold ∈ loss That is |CWCM(θ (t+1) )-CWCM(θ (t) )|<∈ loss, indicating that the loss function has converged; or when the number of iterations reaches the preset maximum value T max , the algorithm stops optimizing further.

[0222] Through an improved weighted average optimization algorithm, the model can adjust key parameters based on the dynamic changes in prediction error, thereby optimizing the LUBE prediction interval. Ultimately, these optimization strategies work together to enable the model to more accurately characterize uncertainty in optical power prediction and improve the reliability of the prediction interval.

[0223] Furthermore, step five includes:

[0224] The K-fold cross-validation method is used to evaluate the optimized interval prediction model. First, the complete dataset is divided into K subsets of equal proportions. Then, in each round of training, K-1 subsets are selected as training sets, and the remaining subset is selected as the test set. The model is trained and the prediction error is calculated. The test set is rotated in sequence, and this is repeated K times so that all data participate in training and validation. Finally, the mean of all K rounds of test errors is used as the final evaluation indicator to obtain the optimal interval prediction model for multi-site optical power prediction.

[0225] The optical power prediction system based on multi-site spatiotemporal characteristics and dynamic optimization includes a data acquisition module, a data screening module, a data fusion module, an interval prediction model, and an optical power prediction module;

[0226] The data acquisition module collects optical power data, cloud image data and meteorological data of multiple monitoring sites;

[0227] The data screening module uses a selection-deletion feature selection method to perform feature selection to screen out the features most relevant to optical power prediction;

[0228] The data fusion module adopts a multi-site data weighted fusion strategy to perform weighted fusion of optical power data, cloud image data and meteorological data according to the error, geographical location and environmental factors of each monitoring site;

[0229] The interval prediction model combines the STG-Mamba-neural network model and the LUBE method; after inputting the feature-selected and weighted fusion data into the STG-Mamba-neural network model, the output is predicted by the LUBE interval to obtain the upper and lower boundaries of the optical power; an improved weighted average optimization algorithm is introduced, combined with random perturbation and boundary constraint strategies, to optimize the interval prediction model;

[0230] The optical power prediction module uses a cross-validation method to evaluate the optimized interval prediction model, and predicts the optical power of multiple sites after obtaining the optimal interval prediction model.

Claims

1. An optical power prediction method based on multi-site spatiotemporal characteristics and dynamic optimization, characterized by: The following steps are involved: Step 1: Collect optical power data, cloud image data, and meteorological data from multiple monitoring stations, and use the select-drop feature selection method to select the features most relevant to optical power prediction; Step 2: Using a multi-site data weighted fusion strategy, optical power data, cloud image data, and meteorological data are weighted and fused according to the error, geographical location, and environmental factors of each monitoring site; Step 3: Construct an interval prediction model that combines the STG-Mamba-neural network model and the LUBE method. Input the feature-selected and weighted fusion data into the STG-Mamba-neural network model and then predict the LUBE interval of the output to obtain the upper and lower boundaries of the optical power. Step 4: Introduce the improved weighted average optimization algorithm, combine random perturbation and boundary constraint strategies, and optimize the interval prediction model; Step 5: Use the cross-validation method to evaluate the optimized interval prediction model, and predict the optical power of multiple sites after obtaining the optimal interval prediction model.

2. The optical power prediction method based on multi-site spatiotemporal characteristics and dynamic optimization according to claim 1 is characterized in that: The select-delete feature selection method comprises the following steps: (1) The importance of features is evaluated using the minimum redundancy and maximum relevance criteria; (2) The exploration strategy in the feature selection process adopts the position update rule of the weighted average optimization algorithm; (3) Selection-deletion operation: By adjusting the probability of deleting features DR and the probability of selecting features SR, a balance is found between global search and local search; The probability of deleting a feature DR and the probability of selecting a feature SR are dynamically adjusted in each iteration: SR=1-DR Among them, DR max and DR min is the maximum and minimum probability of deleting features, MaxFEs is the maximum number of evaluations, and FEs is the current number of evaluations.

3. The optical power prediction method based on multi-site spatiotemporal characteristics and dynamic optimization according to claim 1 is characterized in that: The multi-site data weighted fusion strategy is specifically: considering data quality, geographical location and environmental factors to fuse the optical power data of multiple sites. The formula is as follows: Among them, D 融合 Denotes the fused dataset, D i represents the dataset of the i-th site, w i represents the weight of the i-th site data, and n represents the total number of sites; in: They represent the weights based on data quality, geographical location, and environmental factors, respectively. α, β, and γ are hyperparameters that can be adjusted according to actual needs. The data quality-based weight takes into account the historical errors of each site's data and assigns higher weights to sites with smaller errors. The specific calculation formula is: Among them, MSE i represents the mean square error of site i. Data with smaller mean square error are more reliable, so their corresponding weight is higher. The weights based on geographic location, if there is a strong spatial correlation between sites, the monitoring data located close to the target site should be given a higher weight; the Gaussian kernel function is used to calculate the spatial correlation: Among them, x i and x j denote the geographical coordinates of site i and target site j, σ spatial It is a parameter that controls the degree of influence of spatial distance on correlation, called spatial bandwidth; The calculated K(x i ,x j ) value is directly used as the weight of site i to the target site j data; The weights based on geographic location are normalized so that the sum of the weights of all sites is 1. The normalized weight formula is: The weight based on environmental factors considers the impact of environmental conditions on optical power, including temperature, humidity, and wind speed. The weighting is based on the environmental similarity between the site and the target site. The specific calculation formula is as follows: Among them, T j ,H j ,V j is the temperature, humidity, and wind speed of the target site, T i ,H i ,V i is the temperature, humidity, and wind speed at the i-th station.

4. The optical power prediction method based on multi-site spatiotemporal characteristics and dynamic optimization according to claim 1 is characterized in that: A multi-site data weighted fusion strategy is used to pre-process the data after fusion, including data cleaning and standardization. Data cleaning removes outliers and erroneous records. Normalization is performed using the following formula: Where X is the original data, μ is the mean, σ is the standard deviation, and Z is the standardized data.

5. The optical power prediction method based on multi-site spatiotemporal characteristics and dynamic optimization according to claim 1 is characterized in that: The interval prediction model includes: (1) Inputting the cloud image, time series optical power data and meteorological data after feature selection and weighted fusion into the STG-Mamba model; (2) The predicted value of the optical power output by the STG-Mamba model and its uncertainty interval are used as the input of the LUBE method; (3) Use the LUBE method to perform interval estimation on the optical power prediction results; (4) The indicators of interval coverage loss PICP and interval width loss PIWL are introduced into the LUBE method; and the interval prediction loss function CWCM is proposed, which is as follows: Among them, λ1 and λ2 are weighting coefficients, which are used to adjust the relative importance of PICP and PIWL respectively; η1 and η2 are attenuation factors, which are used to control the influence of PICP and PIWL on the objective function.

6. The optical power prediction method based on multi-site spatiotemporal characteristics and dynamic optimization according to claim 1 is characterized in that: The fourth step introduces an improved weighted average optimization algorithm to optimize the key parameters of the interval prediction model so that the interval prediction loss function CWCM reaches the optimal value, including the following steps: (1) Initialize parameter population In the weighted average optimization algorithm, first initialize N individuals in the population, that is, the parameters of N groups of interval prediction models: Where n represents the nth individual, i.e. the nth set of hyperparameters, k represents the kth hyperparameter, and d represents the total number of optimized hyperparameters; and represents the boundary of the kth hyperparameter, rand is a uniform random number between [0,1], ensuring that the initial population is evenly distributed in the search space; (2) Calculate the objective function For each set of hyperparameters θ n , train the interval prediction model and calculate the loss function CWCM: Among them, λ1 and λ2 are weighting coefficients used to adjust the relative importance of PICP and PIWL respectively; η1 and η2 are attenuation factors used to control the influence of PICP and PIWL on the objective function; Compute the weighted center of the current hyperparameter population: Among them, Fitness(θ i ) represents the fitness of the nth candidate solution, θ center Represents the weighted average position, N sel is the number of candidate solutions selected; (3) Introducing random perturbations and boundary constraints: 3.1 Local Search Among them, w1 and w2 are adaptive weights, randn obeys the standard normal distribution, ∈ is the disturbance factor, ∈0 is the initial disturbance amplitude, T max is the maximum number of iterations; t is the number of iterations; 3.2 Global Search Introduce a perturbation factor for global jumps: Among them, γ0 is the initial perturbation, γ controls the perturbation amplitude, and gradually converges with iteration; 3.3 Boundary Constraints: If the new hyperparameters calculated If the value exceeds the set range, it will be corrected to the nearest boundary value: In addition, the reflection boundary mechanism is combined to further optimize: When the hyperparameter exceeds the boundary, it is reflected to the legal area; (4) The improved weighted average optimization algorithm terminates the optimization when any of the following conditions is met: when the difference between the loss function values of two consecutive iterations is less than the preset convergence threshold ∈ loss That is |CWCM(θ (t+1) )-CWCM(θ (t) )|<∈ loss , indicating that the loss function has converged; or when the number of iterations reaches the preset maximum value T max , the algorithm stops optimizing further.

7. The optical power prediction method based on multi-site spatiotemporal characteristics and dynamic optimization according to claim 1 is characterized in that: Step five includes: The K-fold cross-validation method is used to evaluate the optimized interval prediction model. First, the complete dataset is divided into K subsets of equal proportions. Then, in each round of training, K-1 subsets are selected as training sets, and the remaining subset is selected as the test set. The model is trained and the prediction error is calculated. The test set is rotated in sequence, and this is repeated K times so that all data participate in training and validation. Finally, the mean of all K rounds of test errors is used as the final evaluation indicator to obtain the optimal interval prediction model for multi-site optical power prediction.

8. The optical power prediction system based on multi-site spatiotemporal characteristics and dynamic optimization is characterized by: Including data acquisition module, data screening module, data fusion module, interval prediction model, and optical power prediction module; The data acquisition module collects optical power data, cloud image data and meteorological data of multiple monitoring sites; The data screening module uses a selection-deletion feature selection method to perform feature selection to screen out the features most relevant to optical power prediction; The data fusion module adopts a multi-site data weighted fusion strategy to perform weighted fusion of optical power data, cloud image data and meteorological data according to the error, geographical location and environmental factors of each monitoring site; The interval prediction model combines the STG-Mamba-neural network model and the LUBE method; after inputting the feature-selected and weighted fusion data into the STG-Mamba-neural network model, the output is predicted by the LUBE interval to obtain the upper and lower boundaries of the optical power; an improved weighted average optimization algorithm is introduced, combined with random perturbation and boundary constraint strategies, to optimize the interval prediction model; The optical power prediction module uses a cross-validation method to evaluate the optimized interval prediction model, and predicts the optical power of multiple sites after obtaining the optimal interval prediction model.

9. The optical power prediction system based on multi-site spatiotemporal characteristics and dynamic optimization according to claim 8, characterized in that: The multi-site data weighted fusion strategy is specifically: considering data quality, geographical location and environmental factors to fuse the optical power data of multiple sites. The formula is as follows: Among them, D 融合 Denotes the fused dataset, D i represents the dataset of the i-th site, w i represents the weight of the i-th site data, and n represents the total number of sites; in: They represent the weights based on data quality, geographical location, and environmental factors, respectively. α, β, and γ are hyperparameters that can be adjusted according to actual needs. The data quality-based weight takes into account the historical errors of each site's data and assigns higher weights to sites with smaller errors. The specific calculation formula is: Among them, MSE i represents the mean square error of site i. Data with smaller mean square error are more reliable, so their corresponding weight is higher. The weights based on geographic location, if there is a strong spatial correlation between sites, the monitoring data located close to the target site should be given a higher weight; the Gaussian kernel function is used to calculate the spatial correlation: Among them, x i and x j denote the geographical coordinates of site i and target site j, σ spatial It is a parameter that controls the degree of influence of spatial distance on correlation, called spatial bandwidth; The calculated K(x i ,x j ) value is directly used as the weight of site i to the target site j data; The weights based on geographic location are normalized so that the sum of the weights of all sites is 1. The normalized weight formula is: The weight based on environmental factors considers the impact of environmental conditions on optical power, including temperature, humidity, and wind speed. The weighting is based on the environmental similarity between the site and the target site. The specific calculation formula is as follows: Among them, T j ,H j ,V j is the temperature, humidity, and wind speed of the target site, T i ,H i ,V i is the temperature, humidity, and wind speed at the i-th station.

10. The optical power prediction system based on multi-site spatiotemporal characteristics and dynamic optimization according to claim 8, characterized in that: The interval prediction model includes: (1) Inputting the cloud image, time series optical power data and meteorological data after feature selection and weighted fusion into the STG-Mamba model; (2) The predicted value of the optical power output by the STG-Mamba model and its uncertainty interval are used as the input of the LUBE method; (3) Use the LUBE method to perform interval estimation on the optical power prediction results; (4) The indicators of interval coverage loss PICP and interval width loss PIWL are introduced into the LUBE method; and the interval prediction loss function CWCM is proposed, which is as follows: Among them, λ1 and λ2 are weighting coefficients used to adjust the relative importance of PICP and PIWL respectively; η1 and η2 are attenuation factors used to control the influence of PICP and PIWL on the objective function.

Images