Multivariate photovoltaic power prediction method and device based on automated machine learning

Abstract

This invention discloses a method and apparatus for predicting multivariate photovoltaic power generation based on automated machine learning. The method includes the following steps: Step 1, preparing photovoltaic power generation data; Step 2, performing feature preprocessing and feature selection on the data; Step 3, constructing a model search space, including the model to be searched and its basic structure; Step 4, model optimization, specifically searching the model space and hyperparameters, with each search resulting in a complete model structure and its hyperparameter configuration; Step 5, training the model using training data and evaluating the model, selecting the optimal model and hyperparameters based on the evaluation results, and finally outputting the best trained model; Step 6, predicting multivariate photovoltaic power generation using the best model. This invention can perform data imputation on mixed-type data and achieves good prediction results on photovoltaic datasets with missing attributes and large attribute value fluctuations.

Description

Technical Field

[0001] This invention relates to automated machine learning, specifically a method and apparatus for predicting multivariate photovoltaic power generation based on automated machine learning. Background Technology

[0002] Photovoltaic power generation is an important application technology of solar energy, characterized by randomness and intermittency due to day and night, and is easily affected by weather factors such as solar radiation, temperature, and wind speed. Large-scale photovoltaic power generation connected to the grid can lead to drastic fluctuations in grid voltage, posing significant challenges to the operation and regulation of the power system. Therefore, accurate photovoltaic power generation forecasting enables power system dispatch centers to better plan dispatch schedules and adjust system operating states in a timely manner, reducing energy storage capacity and operating costs, which is of great significance for the stable operation of the power grid system.

[0003] Many artificial intelligence algorithms have been successfully applied in photovoltaic power generation prediction. However, the application of artificial intelligence algorithms in other fields often requires collaboration between domain experts and algorithm engineers, which significantly limits the scope of application and presents certain challenges.

[0004] Automated machine learning is an automated machine learning tool based on neural architecture search. It enables domain experts to automatically build machine learning programs without requiring extensive expertise in artificial intelligence. This is because it can help people with little to no machine learning knowledge build high-quality custom models, automating processes from data feature engineering to model building and hyperparameter selection. It is now widely used in agriculture, industry, and other fields. AutoTS is an automated machine learning framework applied to time series data research. Due to its relatively recent development, no applications have yet been found in photovoltaic power generation prediction. Summary of the Invention

[0005] The purpose of this invention is to provide a method and apparatus for predicting multivariate photovoltaic power generation based on automated machine learning, which can achieve data imputation for mixed-type data and better prediction for photovoltaic data with missing attributes and large fluctuations in attribute values.

[0006] The technical solution adopted in this invention is as follows:

[0007] In a first aspect, the present invention provides a multivariate photovoltaic power generation prediction method based on automated machine learning, which includes the following steps:

[0008] Step 1: Prepare photovoltaic power generation data;

[0009] Step 2: Perform feature preprocessing and feature selection on the data;

[0010] Step 3: Construct the model search space, including the model to be searched and the basic structure of the model;

[0011] Step 4, model optimization, specifically involves searching the model space and hyperparameters. Each search results in a complete model structure and its hyperparameter configuration.

[0012] Step 5: Train the model using the training data, evaluate the model, select the optimal model and hyperparameters based on the evaluation results, and finally output the best model after training.

[0013] Step 6: Predict the multivariate photovoltaic power generation using the optimal model.

[0014] Furthermore, in step 1, the prepared photovoltaic power generation data is divided into three parts in a ratio of 8:1:1: 80% is used as a training set, 10% as a validation set, and 10% as a test set for training.

[0015] Furthermore, in step 2, feature preprocessing includes imputing missing data using the MissForest algorithm. Specifically, the data is first divided into four parts: observed values ​​of the variable, missing values ​​of the variable, observed values ​​other than the variable, and other observed values ​​other than the missing values ​​of the variable. Then, the missing values ​​of the variable are predicted using the mean imputation method. The variables are sorted from smallest to largest by their missing rate. Random forest regression is first used to impute the missing values ​​of the variable with the smallest missing rate. The process is then iterated until the latest imputation result no longer changes from the previous imputation result or the change is very small.

[0016] Furthermore, the iteration index is the magnitude of the change in loss value during iteration, as shown in the following formula:

[0017]

[0018] in, It is an iterative indicator. Represents the current iteration round. This represents the latest filler result. This represents the result of the previous fill.

[0019] Furthermore, in step 2, after preprocessing, feature selection is performed on the multidimensional data including time, irradiance, temperature, humidity, wind direction, wind speed, pressure, actual irradiance, and power generation to filter out redundant and useless features.

[0020] Furthermore, use The scores are evaluated using the following formula:

[0021]

[0022] in, Representing the One characteristic, Representing the The average feature value of positive samples with each feature. Representing the The average feature value of negative samples with each feature. Representing the The average eigenvalue of each feature. Is it for the specific number The first feature For each instance, the denominator is the sum of the variances of the eigenvalues ​​of the positive and negative samples;

[0023] according to The task of predicting photovoltaic power generation is carried out by selecting time, irradiance, and historical power generation as the main features for the fraction size selection.

[0024] Furthermore, the search space is constructed based on three models: Prophet, SARIMAX, and XGBoost, and then the model structure and parameters are optimized, as shown in the following formula:

[0025]

[0026]

[0027] in, This is the optimal neural network model. The optimal learning parameters are... For the neural network search space, and These are the training and validation parts of the input data, respectively, with the cost and cost functions being: The model loss function is ;

[0028] First, optimize the data training part and the neural network search space to find the learning parameters with the lowest loss. Then, use the data validation part and the model with the found optimal learning parameters to find the optimal neural network model based on the cost function.

[0029] Furthermore, simulated annealing is used as the optimization algorithm to automatically select the model structure and hyperparameters. Specifically, the initial solution is calculated first, and the objective function is generated. Then, a new solution is generated by perturbation, and a new objective function is calculated. The change in parameters is obtained by subtracting the two objective function values. If the change is positive, the new solution is accepted according to the Metrapolis criterion; if it is negative, the new solution is accepted directly. Next, it is determined whether the number of iterations has been reached. If not, the previous step is returned. If the number of iterations has been reached, it is determined whether the termination condition is met. If not, the parameter values ​​are slowly reduced and the number of iterations is reset. If the condition is met, the algorithm ends, thus obtaining the optimal parameters.

[0030] Furthermore, the root mean square error (RMSE) is used as the evaluation metric, as shown in the following formula:

[0031]

[0032] in, and These represent the actual value and the predicted value, respectively. and These represent the current round number and the total number of rounds, respectively.

[0033] Secondly, the present invention also provides a multivariate photovoltaic power generation prediction device (system) based on automated machine learning, comprising:

[0034] The acquisition module is used to acquire photovoltaic power generation data for the area to be predicted.

[0035] The prediction processing module is used to input photovoltaic data of the area to be predicted into a pre-trained optimal prediction model and output the photovoltaic power generation of the area.

[0036] Thirdly, the present invention also provides a device for constructing a multivariate photovoltaic power generation prediction model based on automated machine learning, comprising:

[0037] The data processing unit is used for feature preprocessing and feature selection of photovoltaic power generation data;

[0038] Model construction unit, used to construct the model search space, including the model to be searched and the basic structure of the model;

[0039] The model optimization unit is used to search the model space and hyperparameters. Each search results in a complete model structure and its hyperparameter configuration.

[0040] The model training unit is used to train the model using training data.

[0041] The model evaluation unit is used to evaluate the model, select the optimal model and hyperparameters based on the evaluation results, and finally output the best model after training.

[0042] The beneficial effects of this invention are: it employs the missforest algorithm to impute missing data and uses simulated annealing to automatically search various model spaces to select the optimal model and hyperparameters, eliminating the need for manual model building and parameter tuning. This invention can impute mixed-type data and achieves good prediction results on datasets with missing data and high volatility.

[0043] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0044] Figure 1 A flowchart for automated machine learning processing;

[0045] Figure 2 Fill in the flowchart for the missforest algorithm data;

[0046] Figure 3 Generate flowcharts for the model;

[0047] Figure 4 This is a flowchart of the simulated annealing algorithm. Detailed Implementation

[0048] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0049] Automated machine learning consists of three main processes: data preparation, feature engineering, and model generation and evaluation. Model generation includes model selection and hyperparameter optimization, requiring the definition of the model's search space and its optimization methods. The model search space includes traditional machine learning models and neural network architectures. Optimization methods are mainly divided into hyperparameter optimization and model structure optimization. Optimization algorithms include reinforcement learning, evolutionary algorithms, and gradient descent algorithms.

[0050] This invention provides a multivariate photovoltaic power generation prediction method based on automated machine learning. The automated machine learning processing flow of this method is as follows: Figure 1As shown, after inputting photovoltaic power generation data, the data is first preprocessed, including feature standardization and filling in missing data. Then, feature selection is performed using methods such as principal component analysis. After data processing, a search space is constructed, including the model to be searched and its basic structure. An optimization algorithm is then selected to search the model space and hyperparameters. Each search yields a complete model structure and its hyperparameter configuration. The model is then trained using training data, evaluated according to selected evaluation metrics, and the optimal model and hyperparameters are selected based on the evaluation results. Finally, the best trained model and parameters are output.

[0051] The specific implementation steps are as follows:

[0052] First, the prepared photovoltaic power generation data is divided into three parts in a ratio of 8:1:1: 80% is used as the training set, 10% as the validation set, and 10% as the test set for training.

[0053] During the data preprocessing stage, data visualization processing is required for the selected power plant data. First, the changes in radiance, temperature, power generation, wind speed, and other data over time are plotted, revealing significant data loss.

[0054] This invention uses the missforest algorithm to fill in missing data. The algorithm flowchart is as follows: Figure 2 As shown, the data is first divided into four parts: observed values ​​of the variable, missing values ​​of the variable, observed values ​​other than the variable, and other observed values ​​other than the missing values ​​of the variable. Then, the missing values ​​of the variable are predicted using the mean imputation method. The variables are sorted in ascending order of their missing value rate. Random forest regression is first used to impute the missing values ​​of the variable with the lowest missing value rate. This process is then iterated until the latest imputation result no longer changes from the previous imputation result or the change is very small.

[0055] The iteration index is the magnitude of the change in the loss value during iteration, and the formula is as follows: As shown:

[0056] (1)

[0057] in, It is an iterative indicator. Represents the current iteration round. This represents the latest filler result. This represents the result of the previous fill.

[0058] Then, feature selection is performed on multi-dimensional data such as time, irradiance, temperature, humidity, wind direction, wind speed, pressure, actual irradiance, and power generation, filtering out redundant and useless features. This invention uses... The score is used for evaluation, and the formula is as follows: As shown:

[0059] (2)

[0060] in, Representing the One characteristic, Representing the The average feature value of positive samples with each feature. Representing the The average feature value of negative samples with each feature. Representing the The average eigenvalue of each feature. Is it for the specific number The first feature For each instance, the denominator is the sum of the variances of the eigenvalues ​​of the positive and negative samples.

[0061] according to The task of predicting photovoltaic power generation is carried out by selecting time, irradiance, and historical power generation as the main features for the fraction size selection. Then, the search space is constructed.

[0062] This invention uses the root mean square error (RMSE) as the evaluation index, as shown in the formula below. As shown:

[0063] (3)

[0064] in, and These represent the actual value and the predicted value, respectively. and These represent the current round number and the total number of rounds, respectively.

[0065] Model generation flowchart as follows Figure 3 As shown, the search space is constructed based on three models: Prophet, SARIMAX, and XGBoost. Then, the model structure and parameters are optimized, as shown in the formula. As shown:

[0066] (4)

[0067] (5)

[0068] in, This is the optimal neural network model. These are the optimal learning parameters. For the neural network search space, and These are the training and validation parts of the input data, respectively, with the cost and cost functions being: The model loss function is .

[0069] First, optimize the data training part and the neural network search space to find the learning parameters with the lowest loss. Then, use the data validation part and the model with the found optimal learning parameters to find the optimal neural network model based on the cost function.

[0070] This invention uses simulated annealing as an optimization algorithm to automatically select model structure and hyperparameters. The flowchart is as follows: Figure 4 As shown, the simulated annealing algorithm obtains optimal parameters through a simulated annealing process. First, an initial solution is calculated, generating the objective function. Then, a new solution is generated through perturbation, and a new objective function is calculated. The difference between the two objective function values ​​yields the change in parameters. If the change is positive, the new solution is accepted according to the Metrapolis criterion; if negative, the new solution is accepted directly. Next, it checks if the required number of iterations has been reached. If not, it returns to the previous step; if so, it checks if the termination condition is met. If not, the parameter value is slowly reduced, and the iteration count is reset; if the condition is met, the algorithm terminates, thus obtaining the optimal parameters.

[0071] The model was trained using training data, and then evaluated. Based on a comparison of the RMSE of the three models, XGBoost was found to perform best, thus yielding the optimal model. This achieved automated machine learning for predicting multi-dimensional photovoltaic power generation.

[0072] Based on the above method, embodiments of the present invention also provide a multivariate photovoltaic power generation prediction device (system) based on automated machine learning, which includes:

[0073] The acquisition module is used to acquire the photovoltaic dataset for the region to be predicted.

[0074] The prediction processing module is used to input the photovoltaic dataset of the region to be predicted into the pre-trained best prediction model and output the photovoltaic power generation of the region.

[0075] Based on the above method, embodiments of the present invention also provide a device for constructing a multivariate photovoltaic power generation prediction model based on automated machine learning, comprising:

[0076] The data processing unit is used for feature preprocessing and feature selection of photovoltaic power generation data;

[0077] Model construction unit, used to construct the model search space, including the model to be searched and the basic structure of the model;

[0078] The model optimization unit is used to search the model space and hyperparameters. Each search results in a complete model structure and its hyperparameter configuration.

[0079] The model training unit is used to train the model using training data.

[0080] The model evaluation unit is used to evaluate the model, select the optimal model and hyperparameters based on the evaluation results, and finally output the best model after training.

[0081] In summary, this invention facilitates the simple and quick learning of machine learning models for non-experts. It is efficient and cost-effective, employing the missforest algorithm to handle mixed types of missing data and to impute continuous or categorical variables. The simulated annealing algorithm automatically optimizes parameters and model structure. For datasets with missing data and high volatility, this invention demonstrates good predictive performance.

[0082] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the scope of protection of the present invention in any way, and all technical solutions obtained by equivalent substitution or other means fall within the scope of protection of the present invention.

[0083] All parts not covered in this invention are the same as or can be implemented using existing technologies.

Claims

1. A multivariate photovoltaic power generation prediction method based on automated machine learning, characterized in that... Includes the following steps: Step 1: Prepare photovoltaic power generation data; Step 2: Perform feature preprocessing and feature selection on the data; Step 3: Construct the model search space, including the model to be searched and the basic structure of the model; Step 4, model optimization, specifically involves searching the model space and hyperparameters. Each search results in a complete model structure and its hyperparameter configuration. Step 5: Train the model using the training data and evaluate the model. Select the optimal model and hyperparameters based on the root mean square error evaluation results, and finally output the best model after training. include: The search space is constructed based on three models: Prophet, SARIMAX, and XGBoost. Then, the model structure and parameters are optimized, as shown in the following formula: in, This is the optimal neural network model. The optimal learning parameters are... For the neural network search space, and These are the training and validation parts of the input data, respectively, with the cost and cost functions being: The model loss function is ; First, optimize the data training part and the neural network search space to find the learning parameters with the lowest loss. Then, find the optimal model based on the root mean square error of the data validation part and the model with the found learning parameters. Step 6: Predict the multivariate photovoltaic power generation using the optimal model.

2. The method for predicting multivariate photovoltaic power generation based on automated machine learning according to claim 1, characterized in that, In step 1, the prepared photovoltaic power generation data is divided into three parts in a ratio of 8:1:1: 80% is used as the training set, 10% as the validation set, and 10% as the test set for training.

3. The method for predicting multivariate photovoltaic power generation based on automated machine learning according to claim 1, characterized in that, In step 2, feature preprocessing includes imputing missing data using the MissForest algorithm. Specifically, the data is first divided into four parts: observed values ​​of the variable, missing values ​​of the variable, observed values ​​other than the variable, and other observed values ​​other than the missing values ​​of the variable. Then, the missing values ​​of the variable are predicted using the mean imputation method. The variables are sorted from smallest to largest by their missing rate. Random forest regression is first used to impute the missing values ​​of the variable with the smallest missing rate. The process is then iterated until the latest imputation result no longer changes from the previous imputation result.

4. The method for predicting multivariate photovoltaic power generation based on automated machine learning according to claim 3, characterized in that, The iteration index is the magnitude of the change in loss value during iteration, and the formula is shown below: in, It is an iterative indicator. Represents the current iteration round. This represents the latest filler result. This represents the result of the previous fill.

5. The method for predicting multivariate photovoltaic power generation based on automated machine learning according to claim 1, characterized in that, In step 2, after preprocessing, feature selection is performed on the multi-dimensional data, including time, irradiance, temperature, humidity, wind direction, wind speed, pressure, actual irradiance, and power generation, to filter out redundant and useless features.

6. The multivariate photovoltaic power generation prediction method based on automated machine learning according to claim 5, characterized in that, use The scores are evaluated using the following formula: in, Representing the One characteristic, Representing the The average feature value of positive samples with each feature. Representing the The average feature value of negative samples with each feature. Representing the The average eigenvalue of each feature. Is it for the specific number The first feature For each instance, the denominator is the sum of the variances of the eigenvalues ​​of the positive and negative samples; according to The task of predicting photovoltaic power generation is carried out by selecting time, irradiance, and historical power generation as the main features for the fraction size selection.

7. The method for predicting multivariate photovoltaic power generation based on automated machine learning according to claim 1, characterized in that, Simulated annealing algorithm is used as the optimization algorithm to automatically select model structure and hyperparameters. Specifically, the initial solution is calculated and the objective function is generated. Then, the solution is perturbed to generate a new solution and the new objective function is calculated. The change in parameters is obtained by subtracting the two objective function values. If the change is positive, the new solution is accepted according to the Metrapolis criterion. If the value is negative, the new solution is accepted directly. Then, it is checked whether the number of iterations has been reached. If not, the previous step is returned. If the number of iterations has been reached, it is checked whether the termination condition is met. If not, the parameter value is slowly reduced and the number of iterations is reset. If the condition is met, the algorithm ends, thus obtaining the optimal parameters.

8. The method for predicting multivariate photovoltaic power generation based on automated machine learning according to claim 1, characterized in that, The root mean square error (RMSE) is used as the evaluation metric, and the formula is shown below: in, and These represent the actual value and the predicted value, respectively. and These represent the current round number and the total number of rounds, respectively.

Images