Method, device, equipment, medium and product for new energy prediction task processing

Abstract

This invention provides a method, apparatus, device, medium, and product for processing new energy prediction tasks, applied in the field of new energy technology. The method includes: receiving a new energy prediction task and determining the task meta-feature vector of the new energy prediction task; obtaining the model meta-feature vector of each basic regression model in a pre-set base model pool, and generating a joint feature vector based on the model meta-feature vector and the task meta-feature vector; inputting the joint feature vector into a meta-learner to obtain the prediction performance of each basic regression model on the new energy prediction task; determining the target basic regression model based on the prediction performance of each basic regression model on the new energy prediction task, and executing the new energy prediction task. This achieves the ability to predict the prediction performance of each basic regression model on the new energy prediction task through a meta-learner, thereby determining the target basic regression model and eliminating the need to train models one by one to select the target basic regression model for executing the new energy prediction task, reducing resource consumption and improving prediction efficiency.

Description

Technical Field

[0001] This invention relates to the field of new energy technology, and in particular to methods, apparatus, equipment, media and products for new energy forecasting tasks. Background Technology

[0002] Currently, the power generation capacity of new energy power plants, such as photovoltaic power plants and wind farms, is affected by natural factors such as solar radiation and wind speed. Accurate prediction is the key to grid dispatch and absorption.

[0003] In existing technologies, when predicting the power generation of photovoltaic power plants and wind farms, the general approach is to train multiple prediction models one by one and select the optimal model to predict the power generation, or to rely on manual selection of models and parameter tuning to predict the power generation.

[0004] However, when using the above methods, training the model one by one is computationally expensive and resource-intensive. When there are many models or a large amount of data, the training time can take several hours to several days. Furthermore, manually selecting models and adjusting parameters is highly subjective and has poor reproducibility, which means that a new model needs to be selected for each prediction task. Summary of the Invention

[0005] In view of the above problems, methods, apparatus, equipment, media, and products for new energy forecasting tasks are proposed to overcome or at least partially solve the above problems, including: In a first aspect, the present invention provides a method for processing new energy prediction tasks, characterized in that the method includes: Receive new energy prediction tasks and determine the task element feature vector of the new energy prediction tasks; Obtain the model meta-feature vector of each base regression model in the preset base model pool, and generate a joint feature vector based on the model meta-feature vector and the task meta-feature vector; The joint feature vector is input into a pre-trained meta-learner to obtain the prediction performance of each base regression model on the new energy prediction task. Based on the predictive performance of each basic regression model on the new energy prediction task, a target basic regression model is determined, and the new energy prediction task is executed using the target basic regression model.

[0006] Optionally, the meta-learner is trained in the following manner: Construct a pairing space between the basic regression model and the new energy prediction meta-task in the base model pool. Each pairing has a joint feature vector, which is generated based on the model meta-feature vector of the basic regression model and the task meta-feature vector of the new energy prediction meta-task. The performance of some pairs in the pairing space is evaluated to obtain the prediction performance of the evaluated pairs, and a sparse performance matrix is ​​constructed based on the prediction performance of the evaluated pairs; wherein, the sparse performance matrix is ​​used to record the prediction performance of the evaluated pairs, and one row of the sparse performance matrix corresponds to a basic regression model and one column corresponds to a new energy prediction meta-task. Based on the current sparse performance matrix, a Bayesian optimized surrogate model is trained with the evaluated paired joint feature vectors as input and the prediction performance as output. Using the surrogate model, candidate pairs are screened from the unevaluated pairs in the pairing space, the performance of the candidate pairs is evaluated to obtain the predicted performance of the candidate pairs, and the current sparse performance matrix is ​​iteratively updated based on the predicted performance of the candidate pairs. Based on the iteratively updated sparse performance matrix, the proxy model is incrementally trained until the proxy model meets the convergence condition. Using the converged surrogate model, the performance of the remaining unevaluated pairs in the pairing space is predicted to obtain the predicted performance of the remaining unevaluated pairs, and the predicted performance of the remaining unevaluated pairs is used to complete the current sparse performance matrix. The meta-learner is obtained by training using the completed sparse performance matrix.

[0007] Optionally, performance evaluation is performed on a subset of pairings in the pairing space to obtain the predicted performance of the evaluated pairings, and a sparse performance matrix is ​​constructed based on the predicted performance of the evaluated pairings, including: A portion of the pairs are extracted from the pairing space as an initial sample set, and K-fold cross-validation is performed on each pair in the initial sample set to obtain prediction performance; The obtained prediction performance is used as the prediction performance of the evaluated pair to construct a sparse performance matrix.

[0008] Optionally, K-fold cross-validation is performed on each pair in the initial sample set to obtain prediction performance, including: For each pair in the initial sample set, obtain the dataset corresponding to the new energy prediction meta-task in the pair, and divide the dataset into K subsets; One subset is used as the validation set, and the remaining K-1 subsets are used as the training set. The base regression model in the pair is trained on the training set, and the trained base regression model is tested on the validation set to obtain the performance evaluation results. The mean of the K performance evaluation results is used as the predicted performance of the pair.

[0009] Optionally, filtering candidate pairs from unevaluated pairs in the pairing space includes: The joint feature vectors of the unevaluated pairs in the pairing space are input into the surrogate model to obtain the mean and variance of the prediction performance of each unevaluated pair. The acquisition function value is calculated based on the mean and variance of the prediction performance, and candidate pairs are selected from the unevaluated pairs according to the magnitude of the acquisition function value.

[0010] Optionally, the candidate pairings are evaluated for performance to obtain the predicted performance of the candidate pairings, including: K-fold cross-validation is performed on the candidate pairings to obtain prediction performance.

[0011] Optionally, the convergence condition includes any one or more of the following: The number of iterations has reached the preset maximum value; The proportion of evaluated pairs in the sparse performance matrix reaches a preset density threshold. The proxy model's prediction mean square error on the validation set is less than a preset accuracy threshold or the accuracy improvement over multiple consecutive rounds is less than a preset incremental threshold. The maximum acquisition function value of the candidate pair is lower than the preset gain threshold.

[0012] Optionally, before training the meta-learner using the completed sparse performance matrix, the method further includes: The completed sparse performance matrix is ​​post-processed and calibrated; the post-processing calibration includes one or more of the following: row normalization, column normalization, low-rank matrix factorization, and confidence labeling based on prediction variance; The row normalization includes standardizing the prediction performance of each row in the completed sparse performance matrix; the column normalization includes standardizing the prediction performance of each column in the completed sparse performance matrix; the low-rank matrix decomposition includes performing a low-rank approximate decomposition on the completed sparse performance matrix and reconstructing the matrix using the decomposition results; and the confidence labeling based on prediction variance includes labeling the prediction performance in the completed sparse performance matrix with a confidence level based on the variance output by the surrogate model during prediction.

[0013] Optionally, the task element feature vector includes one or more of the following features: mean irradiance, variance of irradiance, mean wind speed, wind speed fluctuation coefficient, temperature statistics, humidity statistics, power output variance, mutual information between irradiance and power, wind speed temporal entropy, data sample size, proportion of missing values, temporal stationarity, and correlation with natural factors. The model metafeature vector includes one or more of the following features: model type encoding, feature processing method, loss function type, fitting optimizer type, number of model parameters, bias term configuration, regularization coefficient, learning rate, number of iterations, polynomial order, and kernel function parameters.

[0014] Optionally, the new energy forecasting task is used to perform any one or more of the following forecasts: Photovoltaic power forecasting, wind power forecasting, and new energy grid-connected load forecasting.

[0015] Secondly, the present invention provides an apparatus for processing new energy prediction tasks, the apparatus comprising: The task receiving module is used to receive new energy prediction tasks and determine the task element feature vector of the new energy prediction task. The joint feature vector generation module is used to obtain the model meta-feature vector of each base regression model in the preset base model pool, and generate a joint feature vector based on the model meta-feature vector and the task meta-feature vector. The performance prediction module is used to input the joint feature vector into a pre-trained meta-learner to obtain the prediction performance of each base regression model on the new energy prediction task. The model selection module is used to determine the target basic regression model based on the prediction performance of each basic regression model on the new energy prediction task, and to use the target basic regression model to execute the new energy prediction task.

[0016] Thirdly, the present invention provides an electronic device including a processor, a memory, and a computer program stored in the memory and capable of running on the processor, wherein the computer program, when executed by the processor, implements the method described above.

[0017] Fourthly, the present invention provides a computer-readable storage medium on which a computer program is stored, wherein the computer program, when executed by a processor, implements the method described above.

[0018] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the method described above.

[0019] The embodiments of the present invention have the following advantages: In this embodiment of the invention, a new energy prediction task is received, and the task meta-feature vector of the new energy prediction task is determined; the model meta-feature vector of each base regression model in the pre-set base model pool is obtained, and a joint feature vector is generated based on the model meta-feature vector and the task meta-feature vector; the joint feature vector is input into a pre-trained meta-learner to obtain the prediction performance of each base regression model on the new energy prediction task; based on the prediction performance of each base regression model on the new energy prediction task, a target base regression model is determined, and the target base regression model is used to execute the new energy prediction task. This achieves the goal of combining the task meta-features corresponding to the new energy prediction task with the task meta-feature vector by generating a joint feature vector. The meta-features of different base regression models are converted into numerical representations of the same dimension. This allows the meta-learner to predict the performance of each base regression model on new energy prediction tasks based on the joint feature vector without actually training the base regression models. Based on the prediction performance, the target base regression model can be determined and used to execute the prediction task. This avoids the computational overhead and training time caused by training models one by one for each new task, as well as the dependence on human experience in model selection. It achieves zero-sample, automated model selection and real-time deployment for new energy prediction tasks, reduces resource consumption for new energy prediction tasks, and improves the prediction efficiency of new energy prediction tasks. Attached Figure Description

[0020] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description of the present invention will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the steps of a method for processing new energy prediction tasks provided in some embodiments of the present invention; Figure 2 This is a flowchart of the steps of a second method for processing new energy prediction tasks provided in some embodiments of the present invention; Figure 3 This is a structural block diagram of a new energy prediction task processing device provided in some embodiments of the present invention. Detailed Implementation

[0022] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0023] Currently, in the process of engineering deployment of new energy forecasting, the power generation of new energy power plants such as photovoltaic and wind power is significantly affected by natural factors such as solar radiation, wind speed, temperature, and humidity, and the data exhibits strong time-series, nonlinearity, volatility, and regional differences.

[0024] When selecting models for new energy forecasting tasks in different regions and time periods, the following two approaches are generally adopted: Option 1 involves training all candidate models one by one on the new dataset and selecting the optimal model based on the performance on the validation set.

[0025] However, this approach leads to high computational complexity and enormous hardware resource consumption. Especially when there are many candidate models (e.g., more than 10) and the new energy dataset is large (e.g., the sample size reaches millions or more), it results in excessive CPU (Central Processing Unit) / GPU (Graphics Processing Unit) computing power utilization and training time reaching several hours to several days, seriously wasting computing resources and making it difficult to meet the requirements of low latency and efficient deployment in engineering scenarios.

[0026] Option 2 involves relying on human experience to select models and adjust parameters.

[0027] However, this approach leads to high subjectivity and poor reproducibility. Differences in understanding of the characteristics of new energy data (such as the nonlinear mapping between irradiance and power, and the characteristics of wind speed fluctuations) among different technical personnel can result in inconsistent model selection results, making it impossible to form a standardized and scalable prediction scheme.

[0028] Secondly, manually selected models are difficult to adapt to new energy prediction scenarios in different regions (such as areas with abundant sunshine and areas with cloudy skies, and wind fields in plains and mountains) and in different seasons. They also cannot utilize the model adaptation experience accumulated in historical new energy prediction tasks. As a result, each new new energy prediction task requires remodeling and reselection, which increases the cost and cycle of engineering deployment, and makes it difficult to guarantee the stability of prediction accuracy.

[0029] In related technologies, when adopting the above two solutions, the lack of knowledge transfer capability means that the "model-natural features-prediction performance" adaptation rules accumulated in historical new energy prediction tasks cannot be effectively reused. Each new new energy prediction task requires modeling and selection to be completed from scratch, resulting in low prediction efficiency and difficulty in meeting the industrial-scale, high-frequency new energy prediction model selection requirements.

[0030] In other words, the relevant technologies cannot solve the problems of insufficient model adaptability and poor generalization in high-dimensional natural feature scenarios of new energy, nor can they fully utilize the natural laws contained in new energy data to improve the reliability of prediction.

[0031] Based on this, this invention proposes to utilize the natural statistical laws of new energy data (such as the temporal variation laws of irradiance and wind speed, and the correlation laws between natural factors and power generation), gradient boosting calculation laws, and high-dimensional natural feature mapping laws, and introduce Bayesian optimization technology to achieve efficient construction of sparse performance matrices.

[0032] This invention fits the response surface of "model-new energy prediction task-performance" using a Gaussian process surrogate model, and intelligently filters high-value samples (these samples can cover new energy tasks with different natural environmental characteristics and different prediction periods) using a data acquisition function, and gradually fills the sparse performance matrix while controlling the consumption of computing resources.

[0033] This invention can shift the computational burden of model selection from each new energy task during runtime to an offline, one-time preparation stage. By constructing a nonlinear mapping relationship between the meta-features of new energy prediction tasks (including natural environment features) and the model meta-feature space and pairing space, it can achieve zero-sample model recommendation for any new energy prediction task without training all candidate models on new energy data. This reduces computational resource consumption and shortens model selection delay, thus adapting to the real-time scheduling requirements of new energy prediction.

[0034] This invention improves model adaptation accuracy and prediction reliability by making full use of natural laws, thereby achieving efficient, automatic, and standardized model selection in new energy prediction scenarios.

[0035] Reference Figure 1 The diagram illustrates a flowchart of a method for processing a new energy prediction task according to some embodiments of the present invention, which may specifically include the following steps: Step 101: Receive the new energy prediction task and determine the task element feature vector of the new energy prediction task.

[0036] In some embodiments, the new energy forecasting task may be a task of forecasting the power generation or load of photovoltaic power plants, wind farms, or new energy grid connection points.

[0037] In practical applications, the corresponding task feature vector can be obtained by analyzing the new energy power plant dataset associated with the new energy prediction task. This dataset can include historical power data and corresponding natural factor data (such as irradiance, wind speed, temperature, and humidity). New energy power plants refer to locations where new energy power generation is carried out, such as photovoltaic power plants and wind farms.

[0038] For example, for the power prediction task of a photovoltaic power plant, statistical quantities such as the mean irradiance, variance of irradiance, and variance of power output, as well as features such as the mutual information between irradiance and power, can be extracted from the historical data of the photovoltaic power plant. By combining the features in the historical data, the task element feature vector of the power prediction task of the photovoltaic power plant can be determined.

[0039] In some embodiments of the present invention, the new energy prediction task is used to perform any one or more of the following predictions: photovoltaic power prediction, wind power prediction, and new energy grid-connected load prediction.

[0040] In practical applications, photovoltaic power prediction can be based on meteorological data such as solar irradiance and temperature to predict the output power of photovoltaic power plants in the future; wind power prediction can be based on data such as wind speed and wind direction to predict the output power of wind farms.

[0041] New energy grid-connected load forecasting can predict the actual power injected into the grid by new energy power plants. This power can be equal to the power generation minus the losses within the plant or the charging and discharging of energy storage. For example, for a photovoltaic power plant equipped with an energy storage system, if the photovoltaic power generation is predicted to be 80 MW, the energy storage system charges at a power of 5 MW, and the plant's power consumption is 1 MW, then the grid-connected load forecast is 74 MW.

[0042] In some embodiments of the present invention, the task element feature vector includes one or more of the following features: mean irradiance, irradiance variance, mean wind speed, wind speed fluctuation coefficient, temperature statistics, humidity statistics, power output variance, mutual information between irradiance and power, wind speed temporal entropy, data sample size, proportion of missing values, temporal stationarity, and correlation with natural factors.

[0043] In some embodiments, the task meta-feature vector may include one or more of statistical features, information-theoretic features, or structural features. Statistical features may include mean irradiance, irradiance variance, mean wind speed, wind speed fluctuation coefficient, temperature statistics, humidity statistics, and power output variance; information-theoretic features may include mutual information between irradiance and power, and wind speed temporal entropy; structural features may include data sample size, percentage of missing values, temporal stationarity, and correlation with natural factors.

[0044] In practical applications, each value in the task element feature vector of the new energy prediction task can be obtained by statistical, information theory, or time series analysis of the historical dataset of new energy power plants associated with the new energy prediction task.

[0045] Optionally, historical irradiance time series can be obtained, and the arithmetic mean of irradiance values ​​at all sampling times can be calculated to obtain the mean irradiance; the average of the squared deviations of irradiance values ​​at each time point from the mean irradiance can be calculated based on the mean irradiance to obtain the irradiance variance; historical wind speed time series can be obtained, and the average of wind speed values ​​at all times can be calculated to obtain the mean wind speed; the standard deviation of the wind speed series can be obtained, and the standard deviation can be divided by the mean wind speed to obtain the wind speed fluctuation coefficient, or the coefficient of variation of wind speed can be calculated to obtain the wind speed fluctuation coefficient.

[0046] Optionally, historical temperature data can be acquired, and one or more of the average, maximum, minimum, or standard deviation can be calculated to obtain temperature statistics; historical relative humidity data can be acquired, and the average or range of variation can be calculated to obtain humidity statistics; historical power time series can be acquired, and the average of the squared deviations of power from the power mean at each time point can be calculated to obtain the power output variance; irradiance series and power series can be acquired, and the mutual information between them can be calculated using the histogram method or kernel density estimation to obtain the mutual information between irradiance and power.

[0047] Optionally, wind speed time series can be obtained, and the complexity measure of the wind speed series can be obtained by calculating sample entropy, approximate entropy or permutation entropy, and the wind speed time series entropy value can be obtained; the total number of samples in the historical dataset can be counted to obtain the data sample size; the number of missing values ​​(such as fields such as irradiance, power, etc. being empty or abnormal) in the statistical dataset can be counted, and the missing value percentage can be obtained by dividing the number of missing values ​​by the total data size.

[0048] Optionally, unit root tests, such as the Augmented Dickey-Fuller test (ADF), can be performed on the power series or natural factor series. The time series stationarity can be determined based on the test statistic or p-value. The correlation coefficients (such as Pearson coefficient or Spearman coefficient) between each natural factor (irradiance, temperature, humidity, wind speed) and power can be calculated, and the mean, maximum value, or combination characteristics of these coefficients can be used to obtain the correlation of natural factors.

[0049] Step 102: Obtain the model meta-feature vector of each base regression model in the preset base model pool, and generate a joint feature vector based on the model meta-feature vector and the task meta-feature vector.

[0050] In practical applications, the model meta-feature vector and the task meta-feature vector can be concatenated to generate a joint feature vector.

[0051] In some embodiments, the base model pool can contain various base models for regression prediction, such as linear models, tree models, neural network models, etc. Each base regression model can correspond to a model meta-feature vector; a higher-dimensional joint feature vector can be obtained by concatenating the model meta-feature vector with the task meta-feature vector.

[0052] In some embodiments of the present invention, the model meta-feature vector includes one or more of the following features: model type encoding, feature processing method, loss function type, fitting optimizer type, number of model parameters, bias term configuration, regularization coefficient, learning rate, number of iterations, polynomial order, and kernel function parameters.

[0053] In practical applications, the model meta-feature vector of a basic regression model can contain features of various algorithmic structures and computational characteristics.

[0054] In some embodiments, the model meta-feature vector may include one or more of structural features or hyperparameter features. Structural features may include model type encoding, feature processing method, loss function type, fitting optimizer type, number of model parameters, and bias term configuration; hyperparameter features may include regularization coefficient, learning rate, number of iterations, polynomial order, and kernel function parameters.

[0055] Among them, the model type encoding can be obtained by numerically encoding the name of the basic regression model (such as one-hot encoding or label encoding); the feature processing method can indicate whether the basic regression model standardizes the input, performs multinomial expansion, etc.; the loss function type can indicate the optimization objective used by the basic regression model (such as mean squared error, absolute error); and the fitting optimizer type can indicate the parameter solving algorithm used by the basic regression model (such as gradient descent, coordinate descent, analytical solution).

[0056] The number of model parameters can be used to count the number of trainable parameters in a basic regression model; the bias term configuration can indicate whether the basic regression model includes an intercept term; the regularization coefficient can represent the strength of L1 or L2 regularization; the learning rate can represent the step size of gradient descent; the number of iterations can represent the maximum number of training rounds of the basic regression model; the polynomial order can represent the highest power of the polynomial regression; and the kernel function parameters can represent the parameter values ​​of the kernel function used in a support vector machine or Gaussian process (such as the bandwidth of the radial basis kernel).

[0057] Step 103: Input the joint feature vector into the pre-trained meta-learner to obtain the prediction performance of each base regression model on the new energy prediction task.

[0058] In practical applications, the generated joint feature vector can be input into a pre-trained meta-learner, which can output the prediction performance of each base regression model on the new energy prediction task.

[0059] In some embodiments, the meta-learner can be a performance prediction model trained offline. The meta-learner can predict the performance of each basis regression model on the new energy prediction task based on the joint feature vector. The prediction performance of each basis regression model on the new energy prediction task can refer to a third evaluation metric value, which is the prediction effect that each basis regression model can achieve on the new energy prediction task predicted by the meta-learner. The third evaluation metric value can include the coefficient of determination (R²), root mean square error (RMSE), or mean absolute error (MAE), etc.

[0060] For example, for a photovoltaic power prediction task, by inputting the joint feature vector of the photovoltaic power prediction task and the support vector regression model into the meta-learner, the meta-learner can output that the prediction performance of the support vector regression model has a determination coefficient of 0.85; by inputting the joint feature vector of the photovoltaic power prediction task and the gated recurrent unit model, a determination coefficient of 0.92 can be obtained.

[0061] Step 104: Based on the prediction performance of each basic regression model on the new energy prediction task, determine the target basic regression model, and use the target basic regression model to execute the new energy prediction task.

[0062] In practical applications, a base regression model can be selected as the target base regression model based on the prediction performance of each base regression model output by the meta-learner on the new energy prediction task, and then the target base regression model can be used to perform the new energy prediction task.

[0063] In some embodiments, prediction performance can reflect the expected performance of different basic regression models in new energy prediction tasks. For example, a higher coefficient of determination indicates better prediction accuracy, and a lower root mean square error indicates smaller error. The best-performing basic regression model can be selected as the target basic regression model based on its prediction performance. The selected target basic regression model can then be used to actually predict the future power generation or grid-connected load of new energy power plants.

[0064] For example, the predictive performance of the meta-learner output support vector regression model has a coefficient of determination of 0.85, while the predictive performance of the gated recurrent unit model has a coefficient of determination of 0.92. The gated recurrent unit model with the highest coefficient of determination can be selected as the target basis regression model, and this target basis regression model can be used to make actual predictions for the newly received photovoltaic power prediction task, outputting the power value for a future period of time.

[0065] In this embodiment of the invention, a new energy prediction task is received, and the task meta-feature vector of the new energy prediction task is determined; the model meta-feature vector of each base regression model in the pre-set base model pool is obtained, and a joint feature vector is generated based on the model meta-feature vector and the task meta-feature vector; the joint feature vector is input into a pre-trained meta-learner to obtain the prediction performance of each base regression model on the new energy prediction task; based on the prediction performance of each base regression model on the new energy prediction task, a target base regression model is determined, and the target base regression model is used to execute the new energy prediction task. This achieves the goal of combining the task meta-features corresponding to the new energy prediction task with the task meta-feature vector by generating a joint feature vector. The meta-features of different base regression models are converted into numerical representations of the same dimension. This allows the meta-learner to predict the performance of each base regression model on new energy prediction tasks based on the joint feature vector without actually training the base regression models. Based on the prediction performance, the target base regression model can be determined and used to execute the prediction task. This avoids the computational overhead and training time caused by training models one by one for each new task, as well as the dependence on human experience in model selection. It achieves zero-sample, automated model selection and real-time deployment for new energy prediction tasks, reduces resource consumption for new energy prediction tasks, and improves the prediction efficiency of new energy prediction tasks.

[0066] In some embodiments of the present invention, the meta-learner is trained in the following manner: Sub-step 1031: Construct a pairing space between the base regression model and the new energy prediction meta-task in the base model pool. Each pairing has a joint feature vector, which is generated based on the model meta-feature vector of the base regression model and the task meta-feature vector of the new energy prediction meta-task.

[0067] In practical applications, the meta-task for new energy prediction can be obtained by collecting existing historical datasets of new energy power plants and labeling them. Each meta-task can correspond to a dataset of new energy power plants and has a clear prediction learning objective (such as predicting the power generation capacity in the future).

[0068] In some embodiments, the new energy prediction meta-task can be constructed from historical operating data of multiple different new energy power plants. Each meta-task can include the input characteristics of the new energy power plant (such as natural factor data such as irradiance, wind speed, temperature, humidity, etc., and historical power data) and the prediction target (such as the power value at a future time).

[0069] For example, we can collect wind speed, wind direction, temperature, and power data from a wind farm in Inner Mongolia over the past three years to define a meta-task: predict wind power output for the next hour based on the wind speed and power data from the previous 6 hours. We can collect data from a photovoltaic power station in Qinghai to define another meta-task: predict photovoltaic power output for the next 2 hours based on irradiance and power data from the previous 4 hours. By combining these two meta-tasks into a meta-task set, we can use it for subsequent pairing space construction and meta-learner training.

[0070] In practical applications, each basic regression model in the base model pool can be combined with each meta-task in the new energy prediction meta-task set to form multiple pairs, and all pairs together constitute the pairing space. Each pair corresponds to a joint feature vector, which can be generated by concatenating the model meta-feature vector of the basic regression model with the task meta-feature vector of the meta-task.

[0071] In some embodiments, the pairing space can be the set of Cartesian products between the basic regression models and the meta-tasks, where the Cartesian product set refers to the set formed by listing all possible combinations. For example, if there are M basic regression models and N new energy prediction meta-tasks, the pairing space between the basic regression models and the meta-tasks is obtained by combining and pairing the M basic regression models and the N meta-tasks one by one, resulting in M×N pairs.

[0072] For example, the base model pool contains two base regression models: Support Vector Regression and Gated Recurrent Unit (GRU). The new energy prediction meta-task set contains two meta-tasks: photovoltaic power plant A and wind farm B. By pairing each base regression model with each meta-task, four pairings can be obtained: (SVR,A), (SVR,B), (GRU,A), and (GRU,B).

[0073] For pairing (SVR, A), the model meta-feature vectors of Support Vector Regression (SVR) (such as model type encoding, regularization coefficients, etc.) can be concatenated with the task meta-feature vectors of photovoltaic power station A (such as mean irradiance, power variance, etc.) to form a joint feature vector.

[0074] For pairing (SVR, B), the model meta-feature vectors of support vector regression (such as model type encoding, regularization coefficients, etc.) can be concatenated with the task meta-feature vectors of wind farm B (such as mean irradiance, power variance, etc.) to form a joint feature vector.

[0075] For pairing (GRU, A), the model meta-feature vectors (such as model type encoding, regularization coefficients, etc.) of the gated recurrent unit (GRU) can be concatenated with the task meta-feature vectors (such as mean irradiance, power variance, etc.) of the photovoltaic power station A to form a joint feature vector.

[0076] For pairing (GRU,B), the model meta-feature vectors (such as model type encoding, regularization coefficients, etc.) of the gated cyclic unit can be concatenated with the task meta-feature vectors (such as mean irradiance, power variance, etc.) of the wind farm B to form a joint feature vector.

[0077] In some embodiments, a heterogeneous base learner set adapted to new energy prediction scenarios can be constructed. This refers to the set of basic regression models under consideration. The regression models selected for the new energy prediction problem include a series of M basic regression models such as GLM (Generalized Linear Model), SVR, GRU, and Transformer, to meet the model adaptation requirements of different new energy prediction scenarios and to adapt to the data fluctuation characteristics caused by natural factors such as solar radiation and wind speed.

[0078] Optionally, a new energy prediction annotation meta-task library can also be constructed. The task library consists of N independent, labeled new energy prediction meta-tasks. Each meta-task corresponds to a new energy power plant dataset and a clear technical learning objective. The learning objective of the meta-task is the performance optimization objective of new energy prediction, where the performance optimization objective is the corresponding evaluation index value (such as the coefficient of determination for regression tasks). (RMSE, MAE, etc.) are used to quantify the actual performance of basic regression models on new energy prediction meta-tasks.

[0079] Optionally, for each new energy prediction meta-task The corresponding new energy power plant dataset, based on the natural statistical laws, environmental impact patterns, and data structure characteristics of new energy data, extracts quantifiable features, including statistical features (mean / variance of irradiance intensity, mean / fluctuation coefficient of wind speed, statistical values ​​of temperature and humidity, and variance of power output), information theory features (mutual information between irradiance and power, and temporal entropy of wind speed), and structural features (data dimension, sample size, proportion of missing values, temporal stationarity, and correlation with natural factors), characterizing the core characteristics of new energy data affected by natural factors. Through normalization and redundancy removal techniques, a K-dimensional task element feature vector is constructed. It is used to characterize the features of the new energy prediction meta-task and its adaptation requirements to the natural environment, providing input for subsequent basic regression model adaptation prediction.

[0080] Optionally, for each base regression model in the base model pool Models such as linear regression, ridge regression, Lasso (Least Absolute Shrinkage and Selection Operator) regression, elastic network regression, and multinomial regression can be used to extract structural element features and hyperparameter features based on the algorithm structure and computational characteristics of the model, combined with the adaptation requirements of new energy prediction scenarios.

[0081] Among them, structural features include model type encoding, feature processing method, loss function type, fitting optimizer type, number of model parameters and bias term configuration, etc., which are used to characterize the structural characteristics of the basic regression model and its computational adaptability to the volatility of new energy data; hyperparameter features include regularization coefficient (adapting to the noise of new energy data), learning rate, number of iterations, polynomial order, kernel function parameters and generalization-related hyperparameters, etc., which are used to characterize the computational performance of the basic regression model and its adaptability to changes in natural factors of new energy.

[0082] Numerical techniques such as one-hot encoding and label encoding are used to process the structural meta-features and hyperparameters. Continuous features retain their original values ​​and are normalized. After being concatenated in a fixed order, they form an L-dimensional model meta-feature vector. .

[0083] In some embodiments, the pairing space between basic regression models in the basic model pool and new energy prediction meta-tasks can be constructed using the following formula, wherein each pairing can be considered as a sample, each sample corresponds to a combination of basic regression models and meta-tasks, and is associated with an evaluation index value of the prediction performance to be evaluated. (like RMSE):

[0084] Where S represents the pairing space, which is the set of all possible combinations of basic regression models and new energy prediction meta-tasks; where, This represents the i-th basic regression model. Let represent the j-th new energy prediction meta-task. The value of i ranges from 1 to M (where M represents the total number of models in the base model pool), and the value of j ranges from 1 to N (where N represents the total number of meta-tasks in the meta-task set).

[0085] Sub-step 1032: Perform performance evaluation on some pairs in the pairing space to obtain the prediction performance of the evaluated pairs, and construct a sparse performance matrix based on the prediction performance of the evaluated pairs; wherein, the sparse performance matrix is ​​used to record the prediction performance of the evaluated pairs, and one row in the sparse performance matrix corresponds to a basic regression model and one column corresponds to a new energy prediction meta-task.

[0086] In practical applications, a subset of pairs can be selected from the pairing space, and the performance of the selected pairs can be evaluated to obtain the predicted performance of the evaluated pairs and construct a sparse performance matrix.

[0087] In some embodiments, performance can be evaluated using K-fold cross-validation, where the predicted performance of the evaluated pair can be a first evaluation metric value of the evaluated pair. The first evaluation metric value is the evaluation metric value obtained through K-fold cross-validation; it can be the coefficient of determination, root mean square error, or mean absolute error, etc.

[0088] In the sparse performance matrix, the positions corresponding to the evaluated pairs can be filled with the corresponding first evaluation index value, and the positions corresponding to the unevaluated pairs can be temporarily marked as missing, such as by using NaN (Not a Number) or blank, thus forming a sparse matrix with most positions empty and a small number of positions having values.

[0089] For example, suppose there are two basic regression models in the basic model pool, such as SVR and GRU, and two meta-tasks in the new energy prediction meta-task set (photovoltaic power plant A and wind farm B).

[0090] By partially pairing pairs in the pairing space, such as (SVR,A), (SVR,B), (GRU,A), and (GRU,B), and specifically (SVR,A) and (GRU,B), performance was evaluated using K-fold cross-validation. The determination coefficients for the first evaluation metric were obtained as 0.85 and 0.92, respectively. A 2x2 sparse performance matrix was constructed, as shown in Table 1.

[0091] Table 1 In some embodiments, for pairing space The Latin hypercube sampling method can be used to extract 5% to 10% of the total sample size to form the initial sample set. This method ensures that the samples cover new energy prediction meta-tasks in different natural environmental scenarios (such as sunny days, cloudy days, windy days, and low temperature days) and different prediction periods (short-term and medium-term). This sampling method can significantly reduce the computational load and resource consumption of the initial sampling while ensuring the representativeness of the samples, avoid the waste of computing power caused by blind sampling, and provide an efficient technical sampling solution for the subsequent construction of the initial performance matrix.

[0092] For the initial sample set Each pair (model-metata task) in the model is paired. Perform K-fold cross-validation to evaluate the predictive stability of the basic regression model under different natural environmental scenarios, and average the results of multiple validations. The first evaluation index value is used as the true generalization performance of the paired basis regression model, and an M×N dimensional initial sparse performance matrix is ​​constructed based on the evaluated first evaluation index value. In the initial sparse performance matrix, the evaluated positions are filled with the first evaluation index value, and the unevaluated positions are marked as missing.

[0093] For each selected pair (basic regression model - meta-task), the steps for performing K-fold cross-validation can be as follows: 1. Divide the new energy power station dataset into K mutually exclusive subsets of similar size.

[0094] 2. Select one subset as the validation set and the remaining K-1 subsets as the training set.

[0095] 3. Train the corresponding basic regression model on the training set, perform performance testing on the validation set, and record the corresponding performance evaluation index results of the basic regression model.

[0096] 4. Repeat K times, so that each subset serves as a validation set exactly once.

[0097] 5. Take the average of the K validation results to obtain the first evaluation index value, which is used as the true performance of the basic regression model on this dataset. Input the corresponding first evaluation index value into the performance matrix.

[0098] By performing K-fold cross-validation on each selected pair, an initial M×N dimensional sparse performance matrix is ​​obtained. .

[0099] In some embodiments, for each evaluated pair (basic regression model - meta-task) sample The task feature vector of the new energy meta-task can be used to... With model meta-feature vector By concatenating and splicing the data, a joint feature vector is obtained. It is used to comprehensively characterize the adaptability of the basic regression model to the new energy prediction meta-task, as well as the adaptability of the basic regression model to natural factors.

[0100] Optionally, the first evaluation index value corresponding to the evaluated pair can be used. Record as a supervision label .

[0101] Sub-step 1033: Based on the current sparse performance matrix, using the evaluated paired joint feature vectors as input and the prediction performance as output, train a Bayesian optimized surrogate model.

[0102] In practical applications, a Bayesian-optimized surrogate model can be trained based on a pre-constructed sparse performance matrix. This model takes the joint feature vectors corresponding to the evaluated pairs as input and the prediction performance of each evaluated pair as output. The Bayesian-optimized surrogate model learns the mapping relationship between the joint feature vectors and the prediction performance.

[0103] In some embodiments, the Bayesian optimized surrogate model can employ a Gaussian Process Regression (GPR) model, which outputs a second evaluation index value and uncertainty variance by combining the joint feature vector of the input. The second evaluation index value is the evaluation index value predicted by the surrogate model; the second evaluation index value can be a coefficient of determination, root mean square error, or mean absolute error, etc.

[0104] For example, the evaluated pairings include (SVR, PV A) and (GRU, PV B), where the joint feature vector of (SVR, PV A) is X1, and the first evaluation metric is a coefficient of determination R² = 0.85; the joint feature vector of (GRU, PV B) is X2, and the first evaluation metric is a coefficient of determination R² = 0.92. By using (X1, 0.85) and (X2, 0.92) as training samples and inputting them into the surrogate model, the surrogate model can learn the performance distribution patterns corresponding to different regions in the joint feature space, thereby providing performance predictions for subsequent unevaluated pairings.

[0105] In some embodiments, joint feature vectors can be used. For input, the actual performance label (i.e., the value of the primary evaluation metric). For output, Gaussian process regression is trained as a surrogate model for Bayesian optimization. The model incorporates the nonlinear distribution patterns of high-dimensional natural feature spaces of new energy sources (such as the nonlinear correlation between irradiance, wind speed, and power). The kernel function of the surrogate model can adopt the Matern52 kernel (a covariance function) and be combined with the ARD (Automatic Relevance Determination) mechanism to adapt to high-dimensional, non-smooth task-model joint feature spaces.

[0106] Optionally, during training, the hyperparameters of the kernel function can be estimated by maximizing the log-marginal likelihood. After training, the surrogate model can output the mean of the corresponding performance predictions (i.e., the second evaluation metric) for any input joint feature vector. With prediction variance .

[0107] Sub-step 1034: Using the surrogate model, candidate pairs are selected from the unevaluated pairs in the pairing space, the performance of the candidate pairs is evaluated to obtain the predicted performance of the candidate pairs, and the current sparse performance matrix is ​​iteratively updated based on the predicted performance of the candidate pairs.

[0108] In practical applications, surrogate models can be used to predict un-evaluated pairs in the pairing space, thereby filtering out candidate pairs worthy of evaluation; performance evaluation can be performed on the candidate pairs to obtain the prediction performance of each candidate pair; based on the prediction performance of the newly obtained candidate pairs, the current sparse performance matrix can be iteratively updated (that is, the prediction performance of the newly evaluated candidate pairs can be filled into the corresponding positions in the matrix).

[0109] In some embodiments, the surrogate model can output the mean of the predicted performance (i.e., the second evaluation index value) and the predicted variance for the joint feature vector of each unevaluated pair. Based on the mean of the predicted performance and the predicted variance, the acquisition function value (such as the upper bound of the expected improvement or the confidence interval of the Gaussian process) can be calculated, and the pair with the higher acquisition function value can be selected as the candidate pair.

[0110] After candidate pairings are determined using the surrogate model, their performance can be evaluated using K-fold cross-validation to obtain the predictive performance of the candidate pairings. The predictive performance of the candidate pairings can be a first evaluation metric value. This first evaluation metric value is obtained through K-fold cross-validation and can be a coefficient of determination, root mean square error, or mean absolute error, etc.

[0111] For example, in the current sparse performance matrix, two pairs, (SVR, PV A) and (GRU, PV B), have been evaluated, while other pairs, such as (SVR, PV B), have not yet been evaluated. The joint eigenvector of (SVR, PV B) can be input into the surrogate model, yielding a determination coefficient of 0.90 and a variance of 0.01 for the second evaluation metric. The calculated acquisition function value is high, therefore it can be selected as a candidate pair. By performing K-fold cross-validation on this candidate pair, the determination coefficient of the first evaluation metric is obtained as 0.89. This 0.89 is then filled into the corresponding position of (SVR, PV B) in the sparse performance matrix.

[0112] In some embodiments, all unevaluated pairings in the pairing space can be traversed, and a joint feature vector corresponding to each unevaluated pair can be constructed by feature concatenation. A candidate point pairing pool is constructed to ensure that the candidate point pairings cover new energy task samples with different natural environment scenarios and different prediction needs.

[0113] Optionally, the distribution characteristics of the high-dimensional joint feature space of new energy sources and the characteristics of sparse samples can be combined. The weighted fusion of Expected Improvement (EI) and Gaussian Process Upper Confidence Bound (GP-UCB) is selected as the combined acquisition function, as shown in the following formula:

[0114] in, This refers to the data acquisition function; This refers to weight; This refers to the expectation of improvement; This refers to the upper bound of the confidence interval for a Gaussian process.

[0115] Optionally, candidate points can be paired and input into the proxy model. Obtain the average performance prediction (i.e., the second evaluation metric value). With prediction variance And calculate the acquisition function value of each candidate point pair. Quantify the evaluation value of each candidate point pair.

[0116] Optionally, the desired improvement in calculation can be obtained based on the currently observed best performance value (i.e., the first evaluation metric value of the evaluated pair) and the performance prediction value of the surrogate model for the unevaluated candidate pair samples (i.e., the second evaluation metric value), as shown in the following formula:

[0117] in, This refers to the best performance value that has been observed so far (i.e., the best evaluation index value among the first evaluation index values ​​that have been evaluated and paired). This refers to the performance prediction value of the surrogate model for unevaluated candidate paired samples (i.e., the second evaluation index value); E represents the mathematical expectation, which is the weighted average of the possible values ​​of the random variable; max represents the maximum value function, which is used to compare two values ​​and return the larger one.

[0118] Optionally, the calculation of the upper bound of the confidence interval for the Gaussian process can introduce an uncertainty weighting coefficient, which can adaptively decay with the number of iterations, as shown in the following formula:

[0119] in, This refers to the uncertainty weighting coefficient, which adaptively decays with the number of iterations.

[0120] Optionally, the acquisition function values ​​of all candidate point pairs can be sorted, and the top k candidate points with higher acquisition function values ​​can be selected as high-value candidate points. This introduces diversity constraints to prevent the selected candidate points from becoming overly clustered in the feature space.

[0121] For example, the cosine distance between each candidate point can be calculated based on the joint feature vectors corresponding to the candidate points. Locally clustered pairs with high similarity (i.e., small cosine distances) can then be eliminated based on this distance, ensuring that the selected candidate points are representative globally and can cover samples from different natural environments. The cosine distance between candidate points can be calculated based on the joint feature vectors corresponding to the candidate points. Calculated.

[0122] Optionally, a K-fold cross-validation process can be performed on each of the selected optimal candidate points in the corresponding new energy prediction meta-task. upper training basis regression model Obtain the true generalization performance (i.e., the first evaluation metric value) for each candidate point. .

[0123] By using these newly obtained observation samples Add to the observed sample set to update the sample set as follows:

[0124] in, This represents the observed sample set at the current iteration t. This represents the updated observed sample set; This represents the joint feature vector corresponding to a candidate pair; This represents the true generalization performance value (first evaluation metric) obtained after performing performance evaluation (i.e., K-fold cross-validation) on the candidate pair.

[0125] After updating the sample set, the initial sparse performance matrix can also be updated simultaneously. , obtained the Sparse matrix of round iteration By using the first evaluation index value of the new assessment Fill in the corresponding matrix The positions that are not evaluated are still marked as missing, thus achieving the gradual filling of the sparse matrix.

[0126] Sub-step 1035: Based on the iteratively updated sparse performance matrix, the surrogate model is incrementally trained until the surrogate model meets the convergence condition.

[0127] After each iteration of updating the sparse performance matrix, the current surrogate model can be incrementally trained based on the newly evaluated pairs contained in the updated matrix until the surrogate model meets the preset convergence condition and the iteration stops.

[0128] In some embodiments, incremental training can be performed by updating or retraining the model parameters using newly added evaluated samples (joint feature vectors and their corresponding first evaluation index values) based on the surrogate model. When any convergence condition is met, it can be determined that the surrogate model has met the convergence condition, and there is no need to continue sampling.

[0129] In some embodiments of the present invention, the convergence conditions include any one or more of the following: the number of iterations reaches a preset maximum value; the proportion of evaluated pairings in the sparse performance matrix reaches a preset density threshold; the mean square error of the prediction of the surrogate model on the validation set is less than a preset accuracy threshold or the accuracy improvement over multiple consecutive rounds is less than a preset increment threshold; the maximum acquisition function value of the candidate pairing is lower than a preset gain threshold.

[0130] In some embodiments, the surrogate model can be incrementally trained based on the updated observed sample set to optimize the hyperparameters of the kernel function and update the fitting results of the performance response surface, thereby providing more accurate predictive support for the next round of iterative sampling. If the convergence condition is not met, the iterative process can be repeated; if the convergence condition is met, sampling can be stopped and the process can proceed to the performance matrix completion stage.

[0131] The convergence condition can be set to stop iteration when any of the following conditions are met: 1. The number of iterations reaches the preset maximum number of iterations. .

[0132] 2. The proportion of evaluated samples in the performance matrix reaches the preset density threshold. .

[0133] 3. The surrogate model's mean squared error on the validation set is less than the preset accuracy threshold. Or, the improvement in prediction accuracy in multiple consecutive iterations is less than the preset incremental threshold. .

[0134] 4. The maximum acquisition function value of the candidate point is lower than the preset gain threshold. This indicates that the exploratory value of the remaining samples has been largely exhausted.

[0135] When any convergence condition is met, the iteration can be stopped and the performance matrix completion stage can be entered; otherwise, iterative sampling can continue.

[0136] Sub-step 1036: Using the converged surrogate model, the performance of the remaining unevaluated pairs in the pairing space is predicted to obtain the predicted performance of the remaining unevaluated pairs, and the predicted performance of the remaining unevaluated pairs is used to complete the current sparse performance matrix.

[0137] In practical applications, once the surrogate model meets the convergence condition, the converged surrogate model can be used to predict the performance of all remaining pairs in the pairing space that have not yet been evaluated, obtain the predicted performance of each remaining unevaluated pair, and fill the corresponding missing positions in the current sparse performance matrix with the predicted performance of the remaining unevaluated pairs, thereby completing the completion of the entire performance matrix.

[0138] In some embodiments, the converged surrogate model has accumulated a sufficient number of real evaluation samples through iterative training, and its performance predictions for unevaluated pairs have high accuracy and reliability. Performance prediction refers to the surrogate model outputting the mean of the predicted performance (i.e., the second evaluation metric) based on the joint feature vector of the input, without needing to perform K-fold cross-validation for performance evaluation.

[0139] In the completed sparse performance matrix, the missing positions can be filled with the second evaluation index value to facilitate subsequent meta-learner training.

[0140] For example, in a pairing space of 100 pairs, after iterative sampling, 30 pairs have been evaluated using K-fold cross-validation, leaving 70 missing positions in the sparse performance matrix. Using a converged Gaussian process surrogate model, the joint feature vectors of these 70 remaining pairs can be input, and the surrogate model outputs a second evaluation index value for each pair. These second evaluation index values ​​are then filled into the 70 missing positions in the matrix, resulting in a complete performance matrix for all 100 pairs. The 30 values ​​represent the first evaluation index value obtained from K-fold cross-validation, and the 70 values ​​represent the second evaluation index value used by the surrogate model for performance prediction.

[0141] In some embodiments, after the iteration converges, the trained surrogate model can be used to predict the performance of all missing positions in the sparse performance matrix, and the second evaluation index value can be filled into the corresponding position to complete the matrix.

[0142] Optionally, the completed sparse performance matrix can be post-processed and calibrated, combining the natural statistical laws of new energy data with noise suppression technology to adapt to power fluctuation noise caused by natural factors.

[0143] Post-processing calibration may include one or more of the following operations: row normalization or column normalization, used to eliminate scale differences between different new energy tasks or different model performance indicators, ensuring the consistency of matrix data; low-rank matrix factorization, used to smooth data noise and improve the reliability of matrix data to adapt to the volatility of new energy data; and confidence labeling based on prediction variance, used to distinguish prediction positions with high, medium and low confidence, especially to highlight prediction positions under extreme natural scenarios, providing a reliability reference for subsequent new energy prediction engineering applications.

[0144] Optionally, a sparse performance matrix containing the first evaluation index value, the second evaluation index value, and the confidence level label can be output. This sparse performance matrix can be used for subsequent meta-learner training to reflect the model's adaptability to different natural environment scenarios, providing a recommended model for the selection of base regression models in new energy prediction tasks.

[0145] Sub-step 1037: The meta-learner is obtained by training using the completed sparse performance matrix.

[0146] In practical applications, the pairs in the completed sparse performance matrix can be used as training data to train a meta-learner. This meta-learner can learn the mapping relationship between the joint feature vector and the prediction performance.

[0147] In some embodiments, the completed sparse performance matrix includes the predictive performance of the basis regression model on the new energy prediction meta-task (including the first evaluation index value and the second evaluation index value). A training sample set can be constructed by using the joint feature vector corresponding to each model-task combination in the sparse performance matrix as input and the corresponding predictive performance as output.

[0148] By training the training sample set using a regression model (such as a lightweight gradient booster), the meta-learner can predict performance on any new joint feature vector. After training, the meta-learner can be used for zero-shot model recommendation for new tasks.

[0149] For example, the completed sparse performance matrix contains 12 paired samples from 3 models and 4 tasks. A lightweight gradient boosting machine model is trained by using these 12 samples as the training set.

[0150] After training is complete, when a new photovoltaic power prediction task is received, the task meta-features of the task and the model meta-features of each model can be concatenated into a joint feature vector, which is then input into the meta-learner to obtain the prediction performance of each model on the new task (i.e., the third evaluation index value), thereby selecting the optimal model.

[0151] In some embodiments, the sparse performance matrix can be based on the completed sparse performance matrix. Construct a training set specifically for meta-learning. For example, each model-task pairing can be used to generate corresponding task-specific meta-feature vectors. With model meta-feature vector Concatenate them into a joint feature vector And the predicted performance value corresponding to the pair in the sparse performance matrix. As regression labels, all pairs in the matrix are iterated through to form a fully covered training sample set. .

[0152] Optionally, LightGBM (Light Gradient Boosting Machine) can be used as the meta-learner for training. It takes the joint feature vector as input and the predicted performance value as output, with the optimization objective of minimizing the prediction error. It learns the non-linear mapping relationship between the task, model, and performance through the gradient boosting mechanism. After training, the meta-learner can be used for zero-shot model performance prediction on new tasks.

[0153] In some embodiments, for the newly added new energy forecasting task in the new energy forecasting scenario It can extract the task meta-feature vector corresponding to the new energy prediction task in real time. And compare it with the model meta-feature vectors of each basic regression model. The features are concatenated separately to form a joint feature vector with the same input dimension as the meta-learner during training. By inputting all these joint feature vectors into a trained lightweight gradient boosting machine learner, the performance value (i.e., the third evaluation metric) of each base regression model on this new energy prediction task can be predicted in batches. The optimal basic regression model is automatically selected based on the ranking of the prediction results (i.e., the values ​​of all third evaluation indicators). This is the recommended model for current new energy forecasting tasks.

[0154] In some embodiments of the present invention, performance evaluation is performed on some pairings in the pairing space to obtain the predicted performance of the evaluated pairings, and a sparse performance matrix is ​​constructed based on the predicted performance of the evaluated pairings, including: A subset of pairs is extracted from the pairing space as an initial sample set, and K-fold cross-validation is performed on each pair in the initial sample set to obtain prediction performance; the obtained prediction performance is used as the prediction performance of the evaluated pairs to construct a sparse performance matrix.

[0155] In practical applications, a portion of the pairs can be extracted from the pairing space as an initial sample set. K-fold cross-validation is then performed on each pair in the initial sample set to obtain the prediction performance of each pair. The predicted performance is then used as the prediction performance of the evaluated pair, and a sparse performance matrix is ​​constructed.

[0156] In some embodiments, partial pairing can be performed using Latin hypercube sampling to ensure uniform coverage of the initial samples across both the model and task dimensions. K-fold cross-validation is a method for evaluating model generalization performance. It divides the dataset into K parts, alternately using one part as the validation set and the rest as the training set, and takes the average of the K validation results as the first evaluation metric. In the sparse performance matrix, the evaluated positions are filled with the true performance value (i.e., the first evaluation metric value), and the unevaluated positions are marked as missing.

[0157] For example, suppose there are 3 basic regression models and 4 new energy prediction meta-tasks in the pairing space, totaling 12 pairings. Latin hypercube sampling is used to extract 20% (approximately 2 to 3) of these pairings as the initial sample set, for example, selecting (SVR, PV A) and (GRU, Wind Farm B). Five-fold cross-validation is performed on (SVR, PV A): the PV A dataset is divided into 5 parts, and 4 parts are used to train SVR and 1 part to test, resulting in 5 error values. The average of these five error values ​​is used as the first evaluation metric for this pairing (e.g., coefficient of determination 0.85). Similarly, five-fold cross-validation is performed on (GRU, Wind Farm B), yielding a coefficient of determination of 0.92. Then, a 3x4 sparse performance matrix is ​​constructed: 0.85 is entered in the PV A column of the SVR row, 0.92 is entered in the Wind Farm B column of the GRU row, and the remaining 10 positions are marked as missing.

[0158] In some embodiments of the present invention, K-fold cross-validation is performed on each pair in the initial sample set to obtain prediction performance, including: For each pair in the initial sample set, obtain the dataset corresponding to the new energy prediction meta-task in the pair, and divide the dataset into K subsets; sequentially use one subset as the validation set and the remaining K-1 subsets as the training set, train the basis regression model in the pair on the training set, and test the trained basis regression model on the validation set to obtain the performance evaluation result; take the mean of the K performance evaluation results as the prediction performance of the pair.

[0159] In practical applications, for each pair in the initial sample set, the dataset of new energy power stations corresponding to the new energy prediction meta-task in the pair can be obtained. Then, the dataset is divided into K non-overlapping subsets. One subset is used as the validation set and the remaining K-1 subsets are used as the training set. The base regression model in the pair is trained on the training set and tested on the validation set to obtain a performance evaluation result. The above process is repeated until each subset has served as a validation set once. The mean of the K performance evaluation results is used as the prediction performance of the pair.

[0160] In some embodiments, K in K-fold cross-validation can be an integer such as 5 or 10. The dataset can be partitioned using random sampling or chronological segmentation to maintain independence between subsets. Performance evaluation results can be quantified using metrics such as coefficient of determination, root mean square error, or mean absolute error. Using the mean value can reduce random fluctuations in a single validation, making the prediction performance more reliable.

[0161] For example, the initial sample set contains one pair (Support Vector Regression model, prediction task for a photovoltaic power station). The dataset for this photovoltaic power station contains 8760 hours of irradiance, temperature, and power data. The dataset is divided into 5 subsets, each containing 1752 samples. First, subset 1 is used as the validation set, and subsets 2-5 are used as the training set to train the Support Vector Regression model and test it on the validation set, yielding a determination coefficient of 0.82 for the evaluation index. Second, subset 2 is used as the validation set, and subsets 1, 3, 4, and 5 are used as the training set, and the model is trained and tested again, yielding a determination coefficient of 0.85 for the evaluation index. Similarly, the third, fourth, and fifth validations are performed, yielding determination coefficients of 0.83, 0.84, and 0.86, respectively. The average of the five determination coefficients is calculated to be 0.84, which is taken as the predictive performance of this pair on the photovoltaic power station task (i.e., the first evaluation index).

[0162] In some embodiments of the present invention, screening candidate pairs from unevaluated pairs in the pairing space includes: The joint feature vectors of unevaluated pairs in the pairing space are input into the surrogate model to obtain the mean and variance of the prediction performance of each unevaluated pair; the acquisition function value is calculated based on the mean and variance of the prediction performance, and candidate pairs are selected from the unevaluated pairs according to the magnitude of the acquisition function value.

[0163] In practical applications, the joint feature vectors of the pairs that have not yet undergone performance evaluation in the pairing space can be input into the surrogate model. The surrogate model outputs the mean prediction performance (i.e., the second evaluation index value) and the prediction variance (i.e., the prediction uncertainty) of each unevaluated pair. Based on the mean prediction performance and the prediction variance, the acquisition function value of each unevaluated pair is calculated, and one or more pairs are selected as candidate pairs from the unevaluated pairs according to the magnitude of the acquisition function value.

[0164] In some embodiments, the surrogate model can be a Gaussian process regression model, which can calculate the mean and variance of the predicted values. The acquisition function can be used to select regions with high predicted means and regions with large predicted variances, thereby proactively selecting the candidate pairs most valuable for optimizing the sparse performance matrix. For example, several pairs with the largest acquisition function values ​​can be selected as candidate pairs.

[0165] For example, there are three unevaluated pairs in the pairing space: pair A has a mean prediction performance of 0.85 and a variance of 0.01; pair B has a mean of 0.82 and a variance of 0.05; and pair C has a mean of 0.80 and a variance of 0.10. The EI values ​​are calculated by expecting to improve the acquisition function: EI(A) = 0.02, EI(B) = 0.08, and EI(C) = 0.12. Based on the magnitude of the acquisition function values, pair C with the highest EI value is selected as a candidate pair. Although its prediction mean is low, its large variance means it has a higher probability of improving the sparse performance matrix.

[0166] In some embodiments of the present invention, performance evaluation of the candidate pairings to obtain the predicted performance of the candidate pairings includes: performing K-fold cross-validation on the candidate pairings to obtain the predicted performance.

[0167] In practical applications, K-fold cross-validation can be performed on the selected candidate pairs to obtain the predictive performance of the candidate pairs.

[0168] In some embodiments, the predictive performance of candidate pairings can be a first evaluation metric value of the candidate pairings. The first evaluation metric value is an evaluation metric value obtained through K-fold cross-validation; the first evaluation metric value can be the coefficient of determination, root mean square error, or mean absolute error, etc.

[0169] For example, the selected candidate pair is (gated recurrent unit model, prediction task of a photovoltaic power station). Five-fold cross-validation is performed on this pair: the photovoltaic power station dataset is randomly divided into 5 parts, each containing 20% ​​of the samples.

[0170] The first time, the first set was used as the validation set, and the second to fifth sets were used as the training set. The gated recurrent unit model was trained and tested on the validation set, and the coefficient of determination of the evaluation index was 0.86. The second time, the second set was used as the validation set, and the rest were used as the training set, and the coefficient of determination of the evaluation index was 0.88. The third time, the coefficient of determination of the evaluation index was 0.85, the fourth time, the coefficient of determination of the evaluation index was 0.87, and the fifth time, the coefficient of determination of the evaluation index was 0.89.

[0171] The average value of the coefficient of determination, calculated by the five evaluation index values, is 0.87, which is taken as the predictive performance of the candidate pair (i.e., the first evaluation index value).

[0172] In some embodiments of the present invention, before training the meta-learner using the completed sparse performance matrix, the method further includes: The completed sparse performance matrix is ​​post-processed and calibrated; the post-processing calibration includes one or more of the following: row normalization, column normalization, low-rank matrix factorization, and confidence labeling based on prediction variance; The row normalization includes standardizing the prediction performance of each row in the completed sparse performance matrix; the column normalization includes standardizing the prediction performance of each column in the completed sparse performance matrix; the low-rank matrix decomposition includes performing a low-rank approximate decomposition on the completed sparse performance matrix and reconstructing the matrix using the decomposition results; and the confidence labeling based on prediction variance includes labeling the prediction performance in the completed sparse performance matrix with a confidence level based on the variance output by the surrogate model during prediction.

[0173] In practical applications, row normalization can be used to standardize the predicted performance (such as the first evaluation index value or the second evaluation index value) of each row in the sparse performance matrix; column normalization can be used to standardize the predicted performance (such as the first evaluation index value or the second evaluation index value) of each column in the sparse performance matrix; low-rank matrix decomposition can be used to perform low-rank approximate decomposition on the sparse performance matrix and reconstruct the sparse performance matrix using the decomposition results; confidence labeling based on prediction variance can be used to assign a confidence level to each predicted performance (such as the first evaluation index value or the second evaluation index value) in the sparse performance matrix according to the variance output by the surrogate model during prediction.

[0174] In some embodiments, row normalization can eliminate the inconsistency in prediction performance range between different basic regression models due to their own scale differences; column normalization can eliminate the performance benchmark differences between different new energy prediction meta-tasks due to different difficulties; low-rank matrix factorization can smooth the random noise introduced in the surrogate model prediction process and improve the reliability of matrix data; confidence labeling can divide the prediction variance into multiple levels such as high, medium and low for subsequent reference.

[0175] For example, the completed performance matrix is ​​a 3x4 matrix, where the rows represent 3 models and the columns represent 4 meta-tasks. The completed performance matrix contains multiple predictive performance metrics, i.e., multiple evaluation metrics. Some are the first evaluation metrics used to assess the performance of the pairings through K-fold cross-validation, while others are the second evaluation metrics used to predict the performance of the pairings through a surrogate model, as shown in Table 2.

[0176] Table 2 When performing row normalization, the mean and standard deviation of all four evaluation indicator values ​​(either the first or second evaluation indicator value) in the first row can be calculated. Subtract the mean from each evaluation indicator value and divide by the standard deviation to ensure the data in that row follows a standard normal distribution. For example, for the SVR row, rounded to two decimal places: the mean is approximately (0.85 + 0.82 + 0.78 + 0.80) / 4, and the standard deviation is approximately 0.03. After normalization: 0.85 → 1.33, 0.82 → 0.33, 0.78 → -1.00, and 0.80 → -0.33. The other rows are processed similarly, resulting in Table 3.

[0177] Table 3 When performing column normalization, the mean and standard deviation of all three evaluation index values ​​in the first column can be calculated. Subtract the mean from each evaluation index value and divide by the standard deviation to ensure that the data in this column follows a standard normal distribution. The other columns are processed similarly, resulting in Table 4.

[0178] Table 4 When performing low-rank matrix decomposition, the target rank can be set to 2. The sparse performance matrix corresponding to Table 4 can be decomposed by methods such as singular value decomposition (SVD), the two largest singular values ​​can be truncated and retained, and then the sparse performance matrix can be reconstructed to remove abnormal fluctuations.

[0179] When assigning confidence level, the confidence level can be determined based on the prediction variance output by the surrogate model when making performance predictions for each pair. For example, pairings with a prediction variance less than 0.01 are assigned high confidence, pairings with a prediction variance between 0.01 and 0.05 are assigned medium confidence, and pairings with a prediction variance greater than 0.05 are assigned low confidence.

[0180] Reference Figure 2 The diagram illustrates a flowchart of another method for processing new energy prediction tasks provided by some embodiments of the present invention, which may specifically include the following steps: Step 201: Construct a pairing space between the basic regression model and the new energy prediction meta-task in the basic model pool. Each pairing has a joint feature vector, which is generated based on the model meta-feature vector of the basic regression model and the task meta-feature vector of the new energy prediction meta-task.

[0181] Step 202: Perform performance evaluation on some pairs in the pairing space to obtain the prediction performance of the evaluated pairs, and construct a sparse performance matrix based on the prediction performance of the evaluated pairs; wherein, the sparse performance matrix is ​​used to record the prediction performance of the evaluated pairs, and one row in the sparse performance matrix corresponds to a basic regression model and one column corresponds to a new energy prediction meta-task.

[0182] Step 203: Based on the current sparse performance matrix, train a Bayesian optimized surrogate model with the evaluated paired joint feature vectors as input and the prediction performance as output.

[0183] Step 204: Using the surrogate model, candidate pairs are selected from the unevaluated pairs in the pairing space, the performance of the candidate pairs is evaluated to obtain the predicted performance of the candidate pairs, and the current sparse performance matrix is ​​iteratively updated based on the predicted performance of the candidate pairs.

[0184] Step 205: Based on the iteratively updated sparse performance matrix, the surrogate model is incrementally trained until the surrogate model meets the convergence condition.

[0185] Step 206: Using the converged surrogate model, the performance of the remaining unevaluated pairs in the pairing space is predicted to obtain the predicted performance of the remaining unevaluated pairs, and the predicted performance of the remaining unevaluated pairs is used to complete the current sparse performance matrix.

[0186] Step 207: Use the completed sparse performance matrix to train and obtain the meta-learner.

[0187] Step 208: Receive the new energy prediction task and determine the task element feature vector of the new energy prediction task.

[0188] Step 209: Obtain the model meta-feature vector of each base regression model in the preset base model pool, and generate a joint feature vector based on the model meta-feature vector and the task meta-feature vector.

[0189] Step 210: Input the joint feature vector into the pre-trained meta-learner to obtain the prediction performance of each base regression model on the new energy prediction task.

[0190] Step 211: Based on the prediction performance of each basic regression model on the new energy prediction task, determine the target basic regression model, and use the target basic regression model to execute the new energy prediction task.

[0191] In this embodiment of the invention, by generating a joint feature vector, the task meta-features corresponding to the new energy prediction task and the model meta-features of different base regression models can be converted into numerical representations of the same dimension. This enables the meta-learner to predict the prediction performance of each base regression model on the new energy prediction task based on the joint feature vector without actually training the base regression model. Based on the prediction performance, the target base regression model can be determined and used to execute the prediction task. This avoids the computational overhead and training time caused by training the model one by one for each new task, as well as the dependence on manual experience in model selection. It achieves zero-sample, automated model selection and real-time deployment for the new energy prediction task, reduces the resource consumption of the new energy prediction task, and improves the prediction efficiency of the new energy prediction task.

[0192] It should be noted that, for the sake of simplicity, the method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments of the present invention are not limited to the described order of actions, because according to the embodiments of the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions involved are not necessarily essential to the embodiments of the present invention.

[0193] Reference Figure 3 The diagram shows a structural schematic of a new energy prediction task processing device provided by some embodiments of the present invention, which may specifically include the following modules: The task receiving module 301 is used to receive new energy prediction tasks and determine the task element feature vector of the new energy prediction task. The joint feature vector generation module 302 is used to obtain the model meta-feature vector of each base regression model in the preset base model pool, and generate a joint feature vector based on the model meta-feature vector and the task meta-feature vector. The performance prediction module 303 is used to input the joint feature vector into a pre-trained meta-learner to obtain the prediction performance of each base regression model on the new energy prediction task. The model selection module 304 is used to determine the target basic regression model based on the prediction performance of each basic regression model on the new energy prediction task, and to use the target basic regression model to execute the new energy prediction task.

[0194] In some embodiments of the present invention, the meta-learner is trained in the following manner: Construct a pairing space between the basic regression model and the new energy prediction meta-task in the base model pool. Each pairing has a joint feature vector, which is generated based on the model meta-feature vector of the basic regression model and the task meta-feature vector of the new energy prediction meta-task. The performance of some pairs in the pairing space is evaluated to obtain the prediction performance of the evaluated pairs, and a sparse performance matrix is ​​constructed based on the prediction performance of the evaluated pairs; wherein, the sparse performance matrix is ​​used to record the prediction performance of the evaluated pairs, and one row of the sparse performance matrix corresponds to a basic regression model and one column corresponds to a new energy prediction meta-task. Based on the current sparse performance matrix, a Bayesian optimized surrogate model is trained with the evaluated paired joint feature vectors as input and the prediction performance as output. Using the surrogate model, candidate pairs are screened from the unevaluated pairs in the pairing space, the performance of the candidate pairs is evaluated to obtain the predicted performance of the candidate pairs, and the current sparse performance matrix is ​​iteratively updated based on the predicted performance of the candidate pairs. Based on the iteratively updated sparse performance matrix, the proxy model is incrementally trained until the proxy model meets the convergence condition. Using the converged surrogate model, the performance of the remaining unevaluated pairs in the pairing space is predicted to obtain the predicted performance of the remaining unevaluated pairs, and the predicted performance of the remaining unevaluated pairs is used to complete the current sparse performance matrix. The meta-learner is obtained by training using the completed sparse performance matrix.

[0195] In some embodiments of the present invention, performance evaluation is performed on some pairings in the pairing space to obtain the predicted performance of the evaluated pairings, and a sparse performance matrix is ​​constructed based on the predicted performance of the evaluated pairings, including: A portion of the pairs are extracted from the pairing space as an initial sample set, and K-fold cross-validation is performed on each pair in the initial sample set to obtain prediction performance; The obtained prediction performance is used as the prediction performance of the evaluated pair to construct a sparse performance matrix.

[0196] In some embodiments of the present invention, K-fold cross-validation is performed on each pair in the initial sample set to obtain prediction performance, including: For each pair in the initial sample set, obtain the dataset corresponding to the new energy prediction meta-task in the pair, and divide the dataset into K subsets; One subset is used as the validation set, and the remaining K-1 subsets are used as the training set. The base regression model in the pair is trained on the training set, and the trained base regression model is tested on the validation set to obtain the performance evaluation results. The mean of the K performance evaluation results is used as the predicted performance of the pair.

[0197] In some embodiments of the present invention, screening candidate pairs from unevaluated pairs in the pairing space includes: The joint feature vectors of the unevaluated pairs in the pairing space are input into the surrogate model to obtain the mean and variance of the prediction performance of each unevaluated pair. The acquisition function value is calculated based on the mean and variance of the prediction performance, and candidate pairs are selected from the unevaluated pairs according to the magnitude of the acquisition function value.

[0198] In some embodiments of the present invention, performance evaluation is performed on the candidate pairings to obtain the predicted performance of the candidate pairings, including: K-fold cross-validation is performed on the candidate pairings to obtain prediction performance.

[0199] In some embodiments of the present invention, the convergence condition includes any one or more of the following: The number of iterations has reached the preset maximum value; The proportion of evaluated pairs in the sparse performance matrix reaches a preset density threshold. The proxy model's prediction mean square error on the validation set is less than a preset accuracy threshold or the accuracy improvement over multiple consecutive rounds is less than a preset incremental threshold. The maximum acquisition function value of the candidate pair is lower than the preset gain threshold.

[0200] In some embodiments of the present invention, before training the meta-learner using the completed sparse performance matrix, the apparatus further includes: The post-processing calibration module is used to perform post-processing calibration on the completed sparse performance matrix; the post-processing calibration includes one or more of the following: row normalization, column normalization, low-rank matrix factorization, and confidence labeling based on prediction variance; The row normalization includes standardizing the prediction performance of each row in the completed sparse performance matrix; the column normalization includes standardizing the prediction performance of each column in the completed sparse performance matrix; the low-rank matrix decomposition includes performing a low-rank approximate decomposition on the completed sparse performance matrix and reconstructing the matrix using the decomposition results; and the confidence labeling based on prediction variance includes labeling the prediction performance in the completed sparse performance matrix with a confidence level based on the variance output by the surrogate model during prediction.

[0201] In some embodiments of the present invention, the task element feature vector includes one or more of the following features: mean irradiance, irradiance variance, mean wind speed, wind speed fluctuation coefficient, temperature statistics, humidity statistics, power output variance, mutual information between irradiance and power, wind speed temporal entropy, data sample size, proportion of missing values, temporal stationarity, and correlation with natural factors. The model metafeature vector includes one or more of the following features: model type encoding, feature processing method, loss function type, fitting optimizer type, number of model parameters, bias term configuration, regularization coefficient, learning rate, number of iterations, polynomial order, and kernel function parameters.

[0202] In some embodiments of the present invention, the new energy prediction task is used to perform any one or more of the following predictions: photovoltaic power prediction, wind power prediction, and new energy grid-connected load prediction.

[0203] Some embodiments of the present invention also provide an electronic device, including a processor, a memory, and a computer program stored in the memory and capable of running on the processor, wherein the computer program, when executed by the processor, implements the method described above.

[0204] Some embodiments of the present invention also provide a computer-readable storage medium on which a computer program is stored, and which, when executed by a processor, implements the method described above.

[0205] Some embodiments of the present invention also provide a computer program product, including a computer program that, when executed by a processor, implements the method described above.

[0206] As the device embodiment is basically similar to the method embodiment, the description is relatively simple, and relevant parts can be found in the description of the method embodiment.

[0207] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0208] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0209] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, embodiments of the present invention can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of the present invention can take the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0210] Embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0211] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0212] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0213] Although preferred embodiments of the present invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present invention.

[0214] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes the aforementioned element.

[0215] The above provides a detailed description of the method, apparatus, equipment, medium, and product for processing new energy prediction tasks. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for processing new energy forecasting tasks, characterized in that, The method includes: Receive new energy prediction tasks and determine the task element feature vector of the new energy prediction tasks; Obtain the model meta-feature vector of each base regression model in the preset base model pool, and generate a joint feature vector based on the model meta-feature vector and the task meta-feature vector; The joint feature vector is input into a pre-trained meta-learner to obtain the prediction performance of each base regression model on the new energy prediction task. Based on the predictive performance of each basic regression model on the new energy prediction task, a target basic regression model is determined, and the new energy prediction task is executed using the target basic regression model.

2. The method according to claim 1, characterized in that, The meta-learner is trained in the following manner: Construct a pairing space between the basic regression model and the new energy prediction meta-task in the base model pool. Each pairing has a joint feature vector, which is generated based on the model meta-feature vector of the basic regression model and the task meta-feature vector of the new energy prediction meta-task. The performance of some pairs in the pairing space is evaluated to obtain the prediction performance of the evaluated pairs, and a sparse performance matrix is ​​constructed based on the prediction performance of the evaluated pairs; wherein, the sparse performance matrix is ​​used to record the prediction performance of the evaluated pairs, and one row of the sparse performance matrix corresponds to a basic regression model and one column corresponds to a new energy prediction meta-task. Based on the current sparse performance matrix, a Bayesian optimized surrogate model is trained with the evaluated paired joint feature vectors as input and the prediction performance as output. Using the surrogate model, candidate pairs are screened from the unevaluated pairs in the pairing space, the performance of the candidate pairs is evaluated to obtain the predicted performance of the candidate pairs, and the current sparse performance matrix is ​​iteratively updated based on the predicted performance of the candidate pairs. Based on the iteratively updated sparse performance matrix, the proxy model is incrementally trained until the proxy model meets the convergence condition. Using the converged surrogate model, the performance of the remaining unevaluated pairs in the pairing space is predicted to obtain the predicted performance of the remaining unevaluated pairs, and the predicted performance of the remaining unevaluated pairs is used to complete the current sparse performance matrix. The meta-learner is obtained by training using the completed sparse performance matrix.

3. The method of claim 2, wherein, The performance of some pairings in the pairing space is evaluated to obtain the predicted performance of the evaluated pairings. Based on the predicted performance of the evaluated pairings, a sparse performance matrix is ​​constructed, including: A portion of the pairs are extracted from the pairing space as an initial sample set, and K-fold cross-validation is performed on each pair in the initial sample set to obtain prediction performance; The obtained prediction performance is used as the prediction performance of the evaluated pair to construct a sparse performance matrix.

4. The method of claim 3, wherein, Perform K-fold cross-validation on each pair in the initial sample set to obtain prediction performance, including: For each pair in the initial sample set, obtain the dataset corresponding to the new energy prediction meta-task in the pair, and divide the dataset into K subsets; One subset is used as the validation set, and the remaining K-1 subsets are used as the training set. The base regression model in the pair is trained on the training set, and the trained base regression model is tested on the validation set to obtain the performance evaluation results. The mean of the K performance evaluation results is used as the predicted performance of the pair.

5. The method of claim 2, wherein, Filtering candidate pairs from unevaluated pairs in the pairing space includes: The joint feature vectors of the unevaluated pairs in the pairing space are input into the surrogate model to obtain the mean and variance of the prediction performance of each unevaluated pair. The acquisition function value is calculated based on the mean and variance of the prediction performance, and candidate pairs are selected from the unevaluated pairs according to the magnitude of the acquisition function value.

6. The method of claim 2, wherein, The performance of the candidate pairings is evaluated to obtain the predicted performance of the candidate pairings, including: K-fold cross-validation is performed on the candidate pairings to obtain prediction performance.

7. The method according to claim 2, characterized in that, The convergence conditions include any one or more of the following: The number of iterations has reached the preset maximum value; The proportion of evaluated pairs in the sparse performance matrix reaches a preset density threshold. The proxy model's prediction mean square error on the validation set is less than a preset accuracy threshold or the accuracy improvement over multiple consecutive rounds is less than a preset incremental threshold. The maximum acquisition function value of the candidate pair is lower than the preset gain threshold.

8. The method according to any one of claims 2 to 7, characterized in that, Before training the meta-learner using the completed sparse performance matrix, the process also includes: The completed sparse performance matrix is ​​post-processed and calibrated; the post-processing calibration includes one or more of the following: row normalization, column normalization, low-rank matrix factorization, and confidence labeling based on prediction variance; The row normalization includes standardizing the prediction performance of each row in the completed sparse performance matrix; the column normalization includes standardizing the prediction performance of each column in the completed sparse performance matrix; the low-rank matrix decomposition includes performing a low-rank approximate decomposition on the completed sparse performance matrix and reconstructing the matrix using the decomposition results; and the confidence labeling based on prediction variance includes labeling the prediction performance in the completed sparse performance matrix with a confidence level based on the variance output by the surrogate model during prediction.

9. The method according to any one of claims 1 to 7, characterized in that, The task element feature vector includes one or more of the following features: mean irradiance, variance of irradiance, mean wind speed, wind speed fluctuation coefficient, temperature statistics, humidity statistics, power output variance, mutual information between irradiance and power, wind speed temporal entropy, data sample size, proportion of missing values, temporal stationarity, and correlation with natural factors. The model metafeature vector includes one or more of the following features: model type encoding, feature processing method, loss function type, fitting optimizer type, number of model parameters, bias term configuration, regularization coefficient, learning rate, number of iterations, polynomial order, and kernel function parameters.

10. The method according to any one of claims 1 to 7, characterized in that, The new energy forecasting task is used to perform one or more of the following forecasts: Photovoltaic power forecasting, wind power forecasting, and new energy grid-connected load forecasting.

11. An apparatus for new energy prediction task processing, characterized in that, The device includes: The task receiving module is used to receive new energy prediction tasks and determine the task element feature vector of the new energy prediction task. The joint feature vector generation module is used to obtain the model meta-feature vector of each base regression model in the preset base model pool, and generate a joint feature vector based on the model meta-feature vector and the task meta-feature vector. The performance prediction module is used to input the joint feature vector into a pre-trained meta-learner to obtain the prediction performance of each base regression model on the new energy prediction task. The model selection module is used to determine the target basic regression model based on the prediction performance of each basic regression model on the new energy prediction task, and to use the target basic regression model to execute the new energy prediction task.

12. An electronic device, comprising: It includes a processor, a memory, and a computer program stored in the memory and capable of running on the processor, wherein the computer program, when executed by the processor, implements the method as described in any one of claims 1 to 10.

13. A computer-readable storage medium, characterized in that, A computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the method as described in any one of claims 1 to 10.

14. A computer program product, characterised in that, It includes a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 10.

Images