A research and development method for supporting power prediction model research and development platform

Abstract

A research and development method for supporting a power prediction model research and development platform, including a template management module, a component management module, a model management module and an online model module; the template management module is used to create and manage model templates, the model management module is used to manage models, the component management module is used to create and manage algorithm components required for models, and the online model module is used to manage models that have been online; in the algorithm component module, algorithm components are newly created and written, and then connected and combined into a model by dragging and dropping, and finally the model is run and the backtest data is checked, the deviation between the measured and predicted results is compared, and the model parameters are adjusted or the algorithm components are changed. Aiming at the power prediction needs in fields such as wind power generation, the present invention can quickly call various algorithms in a visual scene, and various algorithms can be combined and connected by dragging and dropping to obtain different AI models, thereby greatly improving the research and development efficiency.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a research and development method for supporting a power prediction model research and development platform. Background Art

[0002] The process of developing an AI model can be roughly divided into four steps: data collection and annotation, algorithm development, model training and testing, and model online packaging and calling. Each step requires the call of different professional technicians to complete. If there is no unified management platform, the development process will become very complicated and chaotic, affecting the development progress. Therefore, AI model development platforms have emerged, such as "AI Capability Development Platform and Data Processing Method" (Application No. 2019105915851).

[0003] However, the patent focuses more on the versatility of the platform. The purpose is to turn the AI ​​model development process into a standard and standardized business process from the perspective of data flow, thereby improving production efficiency, and does not make targeted optimizations to the algorithm development process. This is because the data in different fields and industries vary greatly, and it is difficult to provide them with a unified algorithm interface. Therefore, it is impossible for developers from different industry backgrounds to easily call algorithms in a visual scenario.

[0004] Fields such as wind power generation are related to both meteorology and energy, and require the development of a deep fusion model of energy and meteorology to predict power. The data called by this AI model has the following two characteristics: (1) the type and category of the data are relatively consistent; (2) the data processing process is relatively similar. Therefore, the algorithm interface of the AI ​​model can be relatively unified. If an AI model development platform for this field can be made, developers in this field can easily call various algorithms in a visual scene, and then obtain different AI models through different combinations and connections. Finally, the optimal AI model can be found through testing and tuning, which can greatly improve R&D efficiency. However, there is currently no such AI model development platform on the market. Summary of the invention

[0005] In response to the above-mentioned problems existing in the prior art, the present application provides a research and development method to support a power prediction model research and development platform. For power prediction needs in fields such as wind power generation, various algorithms can be quickly called in a visual scenario, and various algorithms can be combined and connected by dragging and dropping to obtain different AI models, and the optimal AI model can be found through testing and tuning.

[0006] The technical solution of the present invention is as follows:

[0007] A platform for supporting the development of power prediction models, including a template management module, a component management module, a model management module and an online model module;

[0008] The template management module is used to create and manage model templates; the model templates are used to apply to product instances in batches and quickly generate corresponding models; the product instances are the power prediction model carriers of the platform;

[0009] There are two types of model templates: training model and evaluation model. The training model is used to train model parameters, and the evaluation model is used to evaluate the performance of different training models applied to the same product instance and score the model.

[0010] The template management module has five operations on model templates:

[0011] (1-1) Application: refers to applying the model template to different product instances to generate different models corresponding to the product instances;

[0012] (1-2) Setting: refers to setting up product instances for testing model templates;

[0013] (1-3) Training: refers to training the model template;

[0014] (1-4) Prediction: refers to using the trained template model to perform prediction and generate prediction results;

[0015] (1-5) Copy: refers to copying another identical model template based on this model template;

[0016] (1-6) Delete: means deleting the model template;

[0017] The model management module is used to manage models, including 8 operations:

[0018] (2-1) Training: refers to executing model training and saving the trained model;

[0019] (2-2) Prediction: refers to executing model prediction and generating model prediction results;

[0020] (2-3) Evaluation: refers to performing model evaluation and updating model scores;

[0021] (2-4) Backtesting: refers to backtesting the data for a specified period of time;

[0022] (2-5) Records: refers to checking the records of model training, prediction, evaluation, and backtesting;

[0023] (2-6) Online: refers to mirroring the trained model and putting it online;

[0024] (2-7) Scheduling: refers to the scheduling of configuration training and prediction;

[0025] (2-8) Delete: refers to deleting the model;

[0026] The component management module is used to create and manage algorithm components required for the model; according to different functions, the algorithm components include 11 types, namely, functional types:

[0027] (3-1) Data import: used to import data needed for model training, prediction, and evaluation;

[0028] (3-2) Pre-processing: used to process the imported data. The processing methods include but are not limited to: data cleaning, missing value processing, outlier processing, noise reduction, and filtering;

[0029] (3-3) Feature selection: used to select features that have an important impact on the target variable from the original features;

[0030] (3-4) Feature transformation: used to transform or convert features in the original data set;

[0031] (3-5) Regression algorithm: used to implement the regression algorithm using classical machine learning theory;

[0032] (3-6) Classification algorithm: used to implement the classification algorithm using classical machine learning theory;

[0033] (3-7) Deep learning: used to implement algorithms using deep learning theory;

[0034] (3-8) Ensemble learning: used to perform secondary processing on the prediction results of multiple regression algorithm components or deep algorithm components;

[0035] (3-9) Post-processing: used to perform secondary processing on the prediction results of a single regression algorithm component or a deep algorithm component;

[0036] (3-10) Data merging: used to merge multiple data into one data;

[0037] (3-11) Evaluation: used to score different models under the same product instance;

[0038] The online model module is used to manage the online models, including three operations:

[0039] (4-1) Online: refers to the model being enabled to provide services externally;

[0040] (4-2) Offline: This means that the model stops providing services to the outside world;

[0041] (4-3) Delete: refers to deleting the model of the online model module.

[0042] Furthermore, the steps to create a new algorithm component are as follows:

[0043] S1-1. Enter the name of the algorithm component;

[0044] S1-2, select the functional type of the algorithm component;

[0045] S1-3. Select an algorithm framework; the algorithm framework includes: sklearn, xgboost, tensorflow, keras, torch;

[0046] S1-4, write the program code of the algorithm component;

[0047] S1-5. Save the program code of the algorithm component.

[0048] A model development method comprises the following steps:

[0049] S2-1. Create a custom algorithm component in the algorithm component module; write the program code of the custom algorithm component and save it;

[0050] S2-2. Select appropriate algorithm components in the algorithm component module, connect and combine them by dragging and dropping, obtain a new model and save it;

[0051] S2-3. Find the model in the model management module and perform the online operation;

[0052] S2-4, find the model in the online model module and perform offline operations;

[0053] S2-5. Find the model in the model management module and perform backtesting operations;

[0054] S2-6. Check the backtest data of the model, compare the deviation between the measured and predicted results, adjust the model parameters or change the algorithm components, and then execute S2-3 to S2-6 again;

[0055] S2-7, saving the model as a model template;

[0056] S2-8. Find the model template in the template management module, perform an application operation, and apply the template to the product instance.

[0057] The beneficial technical effects of the present invention are:

[0058] To meet power prediction needs in areas such as wind power generation, various algorithms can be quickly called in a visual scenario. Various algorithms can be combined and connected by dragging and dropping to obtain different AI models, greatly improving R&D efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 This is the page for creating algorithm components;

[0060] Figure 2 This is the page for creating a new model template;

[0061] Figure 3 This is the page for setting up the test product instance of the template;

[0062] Figure 4 Is the page where the template is applied;

[0063] Figure 5 This is the model management page;

[0064] Figure 6 This is the page where the model is launched;

[0065] Figure 7 This is the page for model scheduling configuration;

[0066] Figure 8 It is the page for viewing model records;

[0067] Fig. 9 This is the page for creating new algorithm components;

[0068] Fig.10 This is the page for launching the model module;

[0069] Fig.11 It is a deviation analysis chart of measured and predicted results. DETAILED DESCRIPTION

[0070] The present invention is described in detail below in conjunction with the accompanying drawings and embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0071] like Figure 1 As shown in the menu bar on the left, the embodiment mainly includes the following functional modules:

[0072] (1) Template management module;

[0073] (2) Model management module;

[0074] (3) Component management module;

[0075] (4) Launch model module.

[0076] 1. Template Management Module

[0077] The template management module is used to create and manage AI model (hereinafter referred to as model) templates. A general algorithm model template is pre-built and then applied to different product instances to automatically generate different models for the corresponding product instances, thus achieving the purpose of batch modeling.

[0078] like Figure 2 As shown in the figure, when creating a new model template, you need to select whether the model is a training model or an evaluation model. The training model is used to train model parameters, and the evaluation model is used to evaluate the score of the training model under the same product instance.

[0079] The newly created template will be displayed in the template management module. In the template management module, you can perform 5 operations on the template: apply, set, predict, train, and delete. "Apply" means applying the template to different product instances to generate different models for the corresponding product instances ( Figure 3 ); "Settings" refers to the test product instance of a setting template ( Figure 4 ); "Train" means training the model template and saving the model; "Predict" means using the trained template model to perform prediction and generate prediction results; "Copy" means copying an identical template based on this template; "Delete" means deleting the template.

[0080] 2. Model Management Module

[0081] like Figure 5 As shown, the model management module is a functional module used to manage the power prediction model, which can realize functions such as model training, prediction, evaluation, and online launch.

[0082] In the model management module, you can perform 8 operations on the model: training, prediction, evaluation, backtesting, online, scheduling, and deletion. "Training" means performing model training and saving the trained model; "Prediction" means performing model prediction and generating model prediction results; "Evaluation" means performing model evaluation and updating model scores; "Backtesting" means backtesting data for a specified time period; "Record" means viewing the records of model training, prediction, evaluation, and backtesting ( Figure 8 ); "Online" means to put the trained model online ( Figure 6 ); "Schedule" refers to the configuration of the training and prediction schedule ( Figure 7 ); "delete" means deleting the model.

[0083] 3. Component Management Module

[0084] The component management module is used to create and manage algorithm components required by the model. Figure 1 As shown, algorithm components include the following types:

[0085] (1) Data import: used to import data used for model training, prediction and evaluation;

[0086] (2) Data export: used to export data during model training;

[0087] (3) Pre-processing: used to process the imported data. The processing methods include but are not limited to: data cleaning, missing value processing, outlier processing, noise reduction, and filtering;

[0088] (4) Feature selection: used to select features that have a significant impact on the target variable from the original features;

[0089] (5) Feature conversion: used to convert or transform the features in the original data set;

[0090] (6) Regression algorithm: used to implement the regression algorithm using classical machine learning theory;

[0091] (7) Classification algorithm: used to implement the classification algorithm using classical machine learning theory;

[0092] (8) Deep learning: used to implement algorithms using deep learning theory;

[0093] (9) Ensemble learning: used to perform secondary processing on the prediction results of multiple regression algorithm components or deep algorithm components;

[0094] (10) Post-processing: used to perform secondary processing on the prediction results of a single regression algorithm component or a deep algorithm component;

[0095] (11) Data merging: used to merge multiple data into one data;

[0096] (12) Evaluation: used to calculate the model score under the same product instance.

[0097] There are three types of algorithm components: general algorithms, custom algorithms, and combined algorithms. General algorithms refer to algorithms that are encapsulated in common algorithm packages on the market and can be used directly by algorithm developers, such as svm in the sklearn package; custom algorithms refer to algorithms that are newly developed by algorithm developers or modified based on general algorithms; combined algorithms refer to new algorithms formed by combining two or more general algorithms or custom algorithms.

[0098] like Fig. 9 As shown, the steps to create a new algorithm component are as follows:

[0099] S1-1. Input component name;

[0100] S1-2. Select the component type, which is the 11 algorithm component types listed above;

[0101] S1-3. Select the algorithm framework; the algorithm framework includes: sklearn, xgboost, tensorflow, keras, torch;

[0102] S1-4, write the program code of the algorithm component;

[0103] S1-5. Save the program code of the algorithm component.

[0104] like Figure 1 As shown, in the algorithm component module, different algorithm components can be connected and combined by dragging and dropping to obtain a new model; the completed model is saved in the model management module.

[0105] 4. Online Model Module

[0106] The online model module is used to manage the models that have been online. You can manually launch and de-launch the models. Fig.10 As shown in the figure, three operations can be performed in the online model module: online, offline and delete. "Online" means that the model is enabled to provide external services; "offline" means that the model stops providing external services; "delete" means deleting the model of the online model module.

[0107] The steps to develop a new model based on the embodiment are as follows:

[0108] S2-1, such as Fig. 9 As shown, create a custom algorithm component in the algorithm component module; write the program code of the custom algorithm component and save it;

[0109] S2-2, such as Figure 1 As shown, select appropriate algorithm components in the algorithm component module, connect and combine them by dragging and dropping, obtain a new model and save it;

[0110] S2-3, such as Figure 5 As shown, find the model in the model management module and perform the online operation;

[0111] S2-4, such as Fig.10 As shown, find the model in the online model module and perform offline operations;

[0112] S2-5, such as Figure 5 As shown, find the model in the model management module and perform backtesting operations;

[0113] S2-6, such as Fig.11 As shown, check the backtest data of the model, compare the deviation between the measured and predicted results, adjust the model parameters or change the algorithm components, and then execute S2-3 to S2-6 again;

[0114] S2-7, saving the model as a model template;

[0115] S2-8, such as Figure 3As shown, find the model template in the template management module, perform the application operation, and apply the template to the product instance. Figure 3 , 4 A template was applied to the day-ahead electricity price forecasting project in Shanxi Province.

[0116] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and implementation modes, and can be fully applicable to various fields suitable for the present invention. For those familiar with the art, for those of ordinary skill in the art, various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principle and spirit of the present invention. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to specific details.

Claims

1. A method for developing a platform for supporting power prediction model development, characterized in that: The platform includes a template management module, a component management module, a model management module and an online model module; The template management module is used to create and manage model templates; the model templates are used to apply to product instances in batches and quickly generate corresponding models; the product instances are the power prediction model carriers of the platform; There are two types of model templates: training model and evaluation model. The training model is used to train model parameters, and the evaluation model is used to evaluate the performance of different training models applied to the same product instance and score the model. The template management module has five operations on model templates: (1-1) Application: refers to applying the model template to different product instances to generate different models corresponding to the product instances; (1-2) Setting: refers to setting up product instances for testing model templates; (1-3) Training: refers to training the model template; (1-4) Prediction: refers to using the trained template model to perform prediction and generate prediction results; (1-5) Copy: refers to copying another identical model template based on this model template; (1-6) Delete: means deleting the model template; The model management module is used to manage models, including 8 operations: (2-1) Training: refers to executing model training and saving the trained model; (2-2) Prediction: refers to executing model prediction and generating model prediction results; (2-3) Evaluation: refers to performing model evaluation and updating model scores; (2-4) Backtesting: refers to backtesting the data for a specified period of time; (2-5) Records: refers to checking the records of model training, prediction, evaluation, and backtesting; (2-6) Online: refers to mirroring the trained model and putting it online; (2-7) Scheduling: refers to the scheduling of configuration training and prediction; (2-8) Delete: refers to deleting the model; The component management module is used to create and manage algorithm components required for the model; according to different functions, the algorithm components include 11 types, namely, functional types: (3-1) Data import: used to import data needed for model training, prediction, and evaluation; (3-2) Pre-processing: used to process the imported data. The processing methods include but are not limited to: data cleaning, missing value processing, outlier processing, noise reduction, and filtering; (3-3) Feature selection: used to select features that have an important impact on the target variable from the original features; (3-4) Feature transformation: used to transform or convert features in the original data set; (3-5) Regression algorithm: used to implement the regression algorithm using classical machine learning theory; (3-6) Classification algorithm: used to implement the classification algorithm using classical machine learning theory; (3-7) Deep learning: used to implement algorithms using deep learning theory; (3-8) Ensemble learning: used to perform secondary processing on the prediction results of multiple regression algorithm components or deep algorithm components; (3-9) Post-processing: used to perform secondary processing on the prediction results of a single regression algorithm component or a deep algorithm component; (3-10) Data merging: used to merge multiple data into one data; (3-11) Evaluation: used to score different models under the same product instance; The online model module is used to manage the online models, including three operations: (4-1) Online: refers to the model being enabled to provide services externally; (4-2) Offline: This means that the model stops providing services to the outside world; (4-3) Delete: refers to deleting the model of the online model module; The research and development method comprises the following steps: S2-1. Create a custom algorithm component in the algorithm component module; write the program code of the custom algorithm component and save it; S2-2. Select appropriate algorithm components in the algorithm component module, connect and combine them by dragging and dropping, obtain a new model and save it; S2-3. Find the model in the model management module and perform the online operation; S2-4, find the model in the online model module and perform offline operations; S2-5. Find the model in the model management module and perform backtesting operations; S2-6. Check the backtest data of the model, compare the deviation between the measured and predicted results, adjust the model parameters or change the algorithm components, and then execute S2-3 to S2-6 again; S2-7, saving the model as a model template; S2-8. Find the model template in the template management module, perform an application operation, and apply the template to the product instance.

2. The method for developing a power prediction model development platform according to claim 1, characterized in that: The steps to create a new algorithm component are as follows: S1-1. Enter the name of the algorithm component; S1-2, select the functional type of the algorithm component; S1-3. Select an algorithm framework; the algorithm framework includes: sklearn, xgboost, tensorflow, keras, torch; S1-4, write the program code of the algorithm component; S1-5. Save the program code of the algorithm component.

Images