Automated machine learning-based new energy power cloud-edge collaborative prediction method and system
Through the new energy power cloud-edge collaborative prediction method of automatic machine learning, the edge server and cloud server work together to select and adjust the power prediction model, solving the problems of low precision and lack of new energy power prediction in traditional technology, achieving higher prediction accuracy and cost reduction.
Patent Information
- Application Number
- PCT/CN2023/102756
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-05-16
- Filing Date
- 2023-06-27
- Publication Date
- 2024-11-21
AI Technical Summary
Traditional new energy power prediction systems are relatively backward in massive data processing, feature screening, model generalization and prediction frameworks, resulting in low prediction accuracy, and numerical weather forecast data and new energy power data are easily missing, affecting the prediction accuracy.
Using the new energy power cloud-edge collaborative prediction method with automatic machine learning, the edge server and cloud server work together, select the appropriate power prediction model based on future numerical weather forecast data and historical output power, perform model adjustment and prediction, and adapt to different fields. The station’s weather forecast data is missing and its usage status.
It improves the accuracy and reliability of new energy power prediction, reduces the system construction, operation and maintenance costs, adapts to the differences between different stations, and achieves more accurate output power prediction.
Smart Images

Figure CN2023102756_21112024_PF_FP_ABST
Abstract
Description
Automatic machine learning-based new energy power cloud-edge collaborative prediction method and system Technical Field
[0001] The present application relates to the field of electric power technology, and in particular to a cloud-edge collaborative prediction method and system for new energy power based on automatic machine learning. Background Art
[0002] How to effectively improve the accuracy of renewable energy power generation prediction, reduce the construction and operation costs of renewable energy power prediction systems, and improve the reliability of renewable energy power prediction are technical problems that have long plagued experts and scholars at home and abroad.
[0003] Traditional technologies mainly adopt a model in which each new energy station independently configures a prediction system and reports it to the dispatching agency. The initial investment cost of independent configuration of a single station remains high. In terms of prediction technology, the current technology is relatively backward in terms of massive new energy data processing, feature screening, model generalization and prediction framework, resulting in low prediction accuracy. At the same time, numerical weather prediction data (hereinafter referred to as NWP) and new energy power data are often missing or interrupted. The above objective engineering and technical problems have brought huge challenges to improving prediction accuracy.
[0004] Traditional renewable energy power prediction systems rarely consider the differences in how renewable energy output power is affected by terrain, climate, and power stations. They use a unified model and model parameters to model various power stations. The model has poor adaptability and low overall prediction accuracy.
[0005] Summary of the Invention
[0006] Based on this, it is necessary to provide a new energy power cloud-edge collaborative prediction method and system based on automatic machine learning that can improve the accuracy of site power prediction in response to the above technical problems.
[0007] In a first aspect, the present application provides a cloud-edge collaborative prediction method for new energy power based on automatic machine learning, which is executed by an edge server. The method includes:
[0008] In response to a power forecast demand for a target new energy station, obtaining future numerical weather forecast data of the target new energy station in a future period and historical output power of a historical period corresponding to the future period;
[0009] Based on the lack of future numerical weather forecast data and the amount of historical output power data, a target power prediction model corresponding to the target new energy site is selected from the power prediction model set. Each optional power prediction model in the power prediction model set is trained and issued by the cloud server based on an automated machine learning algorithm.
[0010] Determine the target operating mode of the target power prediction model based on power prediction requirements; wherein the target operating mode includes the prediction business object, prediction type, and prediction time scale;
[0011] According to the target working mode, the target power prediction model is adjusted, and based on the adjusted target power prediction model, the target output power of the target new energy station in the future period is predicted according to the future numerical weather forecast data of the future period and the historical output power of the historical period corresponding to the future period.
[0012] In one embodiment, based on the lack of future numerical weather forecast data and the amount of historical output power data, a target power prediction model corresponding to the target new energy station is selected from the power prediction model set, including:
[0013] If the target new energy station is identified as not a newly built station or a station not undergoing capacity expansion based on the amount of historical output power data, the data missing level of future numerical weather forecast data is determined based on the missing status of future numerical weather forecast data;
[0014] According to the data missing level, a target power prediction model is selected from the optional power prediction models belonging to the target new energy site in the power prediction model set.
[0015] In one embodiment, based on the lack of future numerical weather forecast data and the amount of historical output power data, a target power prediction model corresponding to the target new energy station is selected from the power prediction model set, including:
[0016] If the target new energy station is identified as a newly built station or a station with expanded capacity based on the amount of historical output power data, the data missing level of future numerical weather forecast data is determined based on the missing status of future numerical weather forecast data;
[0017] Based on the basic information of the target new energy station, similar stations whose similarity with the target new energy station reaches a similarity threshold are selected from other new energy stations; wherein, other new energy stations are stations other than the target new energy station in the station set;
[0018] According to the data missing level, the target power prediction model corresponding to the target new energy station is selected from the optional power prediction models belonging to similar stations in the power prediction model set.
[0019] In one embodiment, based on the adjusted target power prediction model, according to future numerical weather forecast data for the future period and historical output power for a historical period corresponding to the future period, the target output power of the target new energy station in the future period is predicted, including:
[0020] If the data missing level is the scattered missing level, the missing data in the future numerical weather forecast data will be supplemented, and based on the adjusted target power prediction model, the target operating mode of the power prediction model corresponding to the target new energy station will be determined according to the historical output power and the supplemented future numerical weather forecast data;
[0021] If the data missing level is a continuous missing level, the target output power of the target new energy station in the future period is predicted based on the adjusted target power prediction model and the historical output power.
[0022] In one embodiment, based on the basic information of the target new energy station, selecting a similar station from other new energy stations whose similarity to the target new energy station reaches a similarity threshold includes:
[0023] Obtaining basic information of the target new energy station and basic information of each other new energy station; wherein the basic information includes at least one dimension of data selected from latitude and longitude data, capacity data, terrain data, landform data, and altitude data;
[0024] Based on the basic information of the target new energy station and the basic information of each other new energy station, determine the similarity value between the target new energy station and each other new energy station in each dimension of data;
[0025] According to the similarity value between the target new energy station and other new energy stations in each dimension of data, similar stations whose similarity with the target new energy station reaches a similarity threshold are selected from other new energy stations.
[0026] In one embodiment, based on the adjusted target power prediction model, according to future numerical weather forecast data for the future period and historical output power for a historical period corresponding to the future period, the target output power of the target new energy station in the future period is predicted, including:
[0027] In one embodiment, if the station type is distributed new energy, the grid location of the target new energy station is obtained;
[0028] Based on the adjusted target power prediction model, the target output power of the target new energy station in the future period is predicted according to the historical output power and future numerical weather forecast data corresponding to the grid location.
[0029] In a second aspect, the present application provides a cloud-edge collaborative prediction method for new energy power based on automatic machine learning, which is executed by a cloud server. The method includes:
[0030] For each new energy station in the station cluster, each initial model is trained based on the training sample set corresponding to the new energy station, and an optional power prediction model corresponding to the new energy station is selected from the trained initial models; the training sample set includes sample weather forecast data and sample output power, and different initial models are constructed based on different algorithms;
[0031] A power prediction model set is constructed based on the optional power prediction models corresponding to each new energy site; wherein the power prediction model set is used for the edge server to select a target power prediction model corresponding to the target new energy site based on the lack of future numerical weather forecast data for the target new energy site in the future period and the amount of data on the historical output power of the target new energy site in the historical period corresponding to the future period; in the target working mode, the target power prediction model is adjusted, and based on the adjusted target power prediction model, the target output power of the target new energy site in the future period is predicted according to the future weather forecast data and the historical output power;
[0032] Send the power prediction model set to the edge server.
[0033] In one embodiment, each initial model is trained based on a training sample set corresponding to the new energy station, including:
[0034] Performing data mining on the original weather data in the original data corresponding to the new energy station to obtain sample weather forecast data in the training sample set;
[0035] Perform multi-stage noise reduction on the original output power in the original data to obtain the sample output power in the training sample set;
[0036] Based on the training sample set, each initial model is trained.
[0037] In one embodiment, data mining is performed on the raw weather data in the raw data corresponding to the new energy station to obtain sample weather forecast data in the training sample set, including:
[0038] Based on slicing technology, the raw weather data in the original data is cleaned;
[0039] Perform correlation analysis on the original output power and the cleaned original weather data to determine the principal component feature set in the original weather data;
[0040] Verify the validity of the principal component feature set to obtain a valid feature set;
[0041] The effective feature set is used as the sample weather forecast data in the training sample set.
[0042] In one embodiment, each initial model is trained based on a training sample set corresponding to the new energy station, including:
[0043] For any initial model, obtain the initial hyperparameter space and initial step size corresponding to the initial model;
[0044] Adjusting the initial step size based on the number of iterations;
[0045] Based on the adjusted step size, search the hyperparameter optimization space within the initial hyperparameter space;
[0046] In the hyperparameter optimization space, based on the initial value of the hyperparameter and the Bayesian optimization algorithm, at least one optimized hyperparameter to be verified is obtained;
[0047] Based on the training sample set corresponding to the new energy station and the super-optimization parameters to be verified, the initial model is subjected to k-fold cross-validation to complete the training of the initial model.
[0048] In one embodiment, selecting an optional power prediction model corresponding to the site from each trained initial model includes:
[0049] Input the verification sample set corresponding to the target new energy station into each trained initial model to obtain the verification output power output by each trained initial model;
[0050] Determine the prediction accuracy of each trained initial model based on the actual output power corresponding to the verification sample in the verification sample set and the verification output power output by each trained initial model;
[0051] According to the prediction accuracy, the optional power prediction model corresponding to the site is selected from the trained initial models.
[0052] In a third aspect, the present application also provides a cloud-edge collaborative prediction device for new energy power based on automatic machine learning, which includes:
[0053] A response module, configured to respond to a power forecast demand for a target new energy station and obtain future numerical weather forecast data of the target new energy station in a future period and a historical output power of a historical period corresponding to the future period;
[0054] A selection module is used to select a target power prediction model corresponding to a target new energy station from a power prediction model set based on the lack of future numerical weather forecast data and the amount of historical output power data. Each selectable power prediction model in the power prediction model set is trained and issued by a cloud server based on an automated machine learning algorithm.
[0055] An adjustment module is used to determine a target operating mode of a target power prediction model according to power prediction requirements; wherein the target operating mode includes a prediction business object, a prediction type, and a prediction time scale;
[0056] The prediction module is used to adjust the target power prediction model according to the target working mode, and based on the adjusted target power prediction model, predict the target output power of the target new energy station in the future period according to the future numerical weather forecast data of the future period and the historical output power of the historical period corresponding to the future period.
[0057] In a fourth aspect, the present application also provides a cloud-edge collaborative prediction device for new energy power based on automatic machine learning, which includes:
[0058] a training module for training each initial model for each new energy station in the station set based on a training sample set corresponding to the new energy station, and selecting an optional power prediction model corresponding to the new energy station from the trained initial models; wherein the training sample set includes sample weather forecast data and sample output power, and different initial models are constructed based on different algorithms;
[0059] A component module is configured to construct a power prediction model set based on the optional power prediction models corresponding to each new energy station; wherein the power prediction model set is used for the edge server to select a target power prediction model corresponding to the target new energy station based on the lack of future numerical weather forecast data of the target new energy station in the future period and the amount of data on the historical output power of the target new energy station in the historical period corresponding to the future period; in the target working mode, the target power prediction model is adjusted, and based on the adjusted target power prediction model, the target output power of the target new energy station in the future period is predicted according to the future weather forecast data and the historical output power;
[0060] The sending module is used to send the power prediction model set to the edge server.
[0061] In a fifth aspect, the present application further provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:
[0062] In response to a power forecast demand for a target new energy station, obtaining future numerical weather forecast data of the target new energy station in a future period and historical output power of a historical period corresponding to the future period;
[0063] Based on the lack of future numerical weather forecast data and the amount of historical output power data, a target power prediction model corresponding to the target new energy site is selected from the power prediction model set. Each optional power prediction model in the power prediction model set is trained and issued by the cloud server based on an automated machine learning algorithm.
[0064] Determine the target operating mode of the target power prediction model based on power prediction requirements; wherein the target operating mode includes the prediction business object, prediction type, and prediction time scale;
[0065] According to the target working mode, the target power prediction model is adjusted, and based on the adjusted target power prediction model, the target output power of the target new energy station in the future period is predicted according to the future numerical weather forecast data of the future period and the historical output power of the historical period corresponding to the future period.
[0066] In a sixth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the following steps:
[0067] In response to a power forecast demand for a target new energy station, obtaining future numerical weather forecast data of the target new energy station in a future period and historical output power of a historical period corresponding to the future period;
[0068] Based on the lack of future numerical weather forecast data and the amount of historical output power data, a target power prediction model corresponding to the target new energy site is selected from the power prediction model set. Each optional power prediction model in the power prediction model set is trained and issued by the cloud server based on an automated machine learning algorithm.
[0069] Determine the target operating mode of the target power prediction model based on power prediction requirements; wherein the target operating mode includes the prediction business object, prediction type, and prediction time scale;
[0070] According to the target working mode, the target power prediction model is adjusted, and based on the adjusted target power prediction model, the target output power of the target new energy station in the future period is predicted according to the future numerical weather forecast data of the future period and the historical output power of the historical period corresponding to the future period.
[0071] In a seventh aspect, the present application further provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:
[0072] For each new energy station in the station cluster, each initial model is trained based on the training sample set corresponding to the new energy station, and an optional power prediction model corresponding to the new energy station is selected from the trained initial models; the training sample set includes sample weather forecast data and sample output power, and different initial models are constructed based on different algorithms;
[0073] A power prediction model set is constructed based on the optional power prediction models corresponding to each new energy site; wherein the power prediction model set is used for the edge server to select a target power prediction model corresponding to the target new energy site based on the lack of future numerical weather forecast data for the target new energy site in the future period and the amount of data on the historical output power of the target new energy site in the historical period corresponding to the future period; in the target working mode, the target power prediction model is adjusted, and based on the adjusted target power prediction model, the target output power of the target new energy site in the future period is predicted according to the future weather forecast data and the historical output power;
[0074] Send the power prediction model set to the edge server.
[0075] In an eighth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the following steps:
[0076] For each new energy station in the station cluster, each initial model is trained based on the training sample set corresponding to the new energy station, and an optional power prediction model corresponding to the new energy station is selected from the trained initial models; the training sample set includes sample weather forecast data and sample output power, and different initial models are constructed based on different algorithms;
[0077] A power prediction model set is constructed based on the optional power prediction models corresponding to each new energy site; wherein the power prediction model set is used for the edge server to select a target power prediction model corresponding to the target new energy site based on the lack of future numerical weather forecast data for the target new energy site in the future period and the amount of data on the historical output power of the target new energy site in the historical period corresponding to the future period; in the target working mode, the target power prediction model is adjusted, and based on the adjusted target power prediction model, the target output power of the target new energy site in the future period is predicted according to the future weather forecast data and the historical output power;
[0078] Send the power prediction model set to the edge server.
[0079] In a ninth aspect, the present application also provides a cloud-edge collaborative prediction system for new energy power based on automatic machine learning, comprising: a cloud server and an edge server;
[0080] The cloud server is configured to train each initial model for each new energy station in the station cluster based on a training sample set corresponding to the new energy station, and select an optional power prediction model corresponding to the new energy station from the trained initial models; wherein the training sample set includes sample weather forecast data and sample output power, and different initial models are constructed based on different algorithms; construct a power prediction model set based on the optional power prediction models corresponding to each new energy station; and distribute the power prediction model set to the edge server;
[0081] An edge server is used to receive a power prediction model set; in response to a power prediction demand for a target new energy station, obtain future numerical weather forecast data of the target new energy station in a future period and historical output power of a historical period corresponding to the future period; based on the missing situation of future numerical weather forecast data and the amount of historical output power data, select a target power prediction model corresponding to the target new energy station from the power prediction model set; wherein, each optional power prediction model in the power prediction model set is trained and issued by a cloud server based on an automatic machine learning algorithm; according to the power prediction demand, determine a target working mode of the target power prediction model; wherein, the target working mode includes a prediction business object, a prediction type and a prediction time scale; according to the target working mode, adjust the target power prediction model, and based on the adjusted target power prediction model, predict the target output power of the target new energy station in the future period according to the future numerical weather forecast data in the future period and the historical output power of the historical period corresponding to the future period.
[0082] In the above-mentioned automatic machine learning new energy power cloud-edge collaborative prediction method and system, in the new energy system, the usage status of each station (reflected by the amount of historical output power data) is different, and the same station uses different future numerical weather forecast data for power prediction in different time periods. The unified power prediction model used in traditional technology cannot adapt to different stations; in this application, based on the station usage status and future numerical weather forecast data missing conditions corresponding to the target new energy station, the target power prediction model corresponding to the target new energy station is accurately matched from the power prediction model set. The target power prediction model can perform power prediction while adapting to the corresponding weather forecast data missing conditions or station usage status, so that accurate prediction of the output power of different stations can be achieved; further, based on the power prediction requirements of the target new energy station, the working mode of the target power prediction model is adjusted to obtain a target output power that is more in line with the prediction requirements, thereby improving the accuracy of power prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0083] FIG1 is a diagram illustrating an application environment of a cloud-edge collaborative prediction method for new energy power using automated machine learning in one embodiment;
[0084] FIG2 is a flow chart of a method for cloud-edge collaborative prediction of new energy power using automated machine learning in one embodiment;
[0085] FIG3 is a schematic diagram of matching corresponding prediction models for stations in different scenarios in one embodiment;
[0086] FIG4 is a schematic diagram of a process for selecting a target power prediction model according to a data missing level when the target station is a newly built station or a newly expanded station in one embodiment;
[0087] FIG5 is a schematic diagram of a process for determining a target new energy station corresponding to a similar station in one embodiment;
[0088] FIG6 is a flowchart illustrating selecting a target power prediction model based on a data missing level when the target station is not a newly built station or a non-expanded station in one embodiment;
[0089] FIG7 is a schematic diagram of completing short-term interrupted future numerical weather forecast data in one embodiment;
[0090] FIG8 is a schematic diagram of an algorithm for constructing a non-NWP model in one embodiment;
[0091] FIG9 is a schematic diagram of an intelligent spare tire mode according to an embodiment;
[0092] FIG10 is a flow chart of a method for cloud-edge collaborative prediction of new energy power using automatic machine learning in another embodiment;
[0093] FIG11 is a schematic diagram of a process for obtaining a training sample set in one embodiment;
[0094] FIG12 is a schematic diagram of a process for obtaining sample weather forecast data in a training sample set according to one embodiment;
[0095] FIG13 is a schematic diagram of data cleaning, multi-stage noise reduction, and meteorological cross-recursive optimization in one embodiment;
[0096] FIG14 is a schematic diagram of a process for training an initial model in one embodiment;
[0097] FIG15 is a schematic diagram of ensemble learning of various models in one embodiment;
[0098] FIG16 is a schematic diagram of a process for determining an optional prediction power model in one embodiment;
[0099] FIG17 is a schematic diagram of performing timed updates on various power prediction models in one embodiment;
[0100] FIG18 is a schematic diagram of the algorithm architecture of full-cycle automatic machine learning in one embodiment;
[0101] Figure 19 is a schematic diagram of the deployment process of the edge-cloud collaborative system prediction model;
[0102] Figure 20 is a schematic diagram of the system framework for edge-cloud collaboration;
[0103] FIG21 is a block diagram of a new energy power cloud-edge collaborative prediction device based on automatic machine learning in one embodiment;
[0104] FIG22 is a block diagram of a new energy power cloud-edge collaborative prediction device based on automatic machine learning in one embodiment;
[0105] FIG23 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION
[0106] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.
[0107] As shown in Figure 1, this embodiment also provides an automatic machine learning new energy power cloud-edge collaborative prediction system, including: a cloud server 104 and an edge server 102, and the edge server 102 communicates with the cloud server 104.
[0108] In one embodiment, as shown in FIG2 , a cloud-edge collaborative prediction method for renewable energy power based on automatic machine learning is provided. The method is described by applying it to the edge server 102 in FIG1 as an example, including the following steps:
[0109] S201 , in response to a power prediction demand for a target new energy station, obtaining future numerical weather forecast data of the target new energy station in a future period and historical output power of a historical period corresponding to the future period.
[0110] The power forecast demand includes information about the target renewable energy station, such as its number, longitude and latitude, capacity, terrain, topography, and altitude. It is understood that the target renewable energy station can be either a master station or a substation in the renewable energy system. It is also understood that the master station is the dispatching center's main station, responsible for dispatching the substations.
[0111] Among them, the future numerical weather forecast data is the numerical weather forecast (NWP) for the future time period. The future numerical weather forecast data can be indicators related to the natural environment in the area where the target new energy station is located. These indicators may include at least one of temperature, cloud cover, shortwave radiation, longwave radiation, surface air pressure, large-scale precipitation, convective precipitation and humidity.
[0112] For example, the future time period is: XX / XX / X3 [09:00-13:00], and accordingly, the historical output power of the historical time periods corresponding to the future time period is: the historical output power of XX / XX / X0 [09:00-13:00], the historical output power of XX / XX / X1 [09:00-13:00], the historical output power of XX / XX / X2 [09:00-13:00], and so on.
[0113] S202 , based on the missing of future numerical weather forecast data and the amount of historical output power data, select a target power prediction model corresponding to the target new energy station from the power prediction model set.
[0114] Among them, each optional power prediction model in the power prediction model set is automatically trained and automatically issued by the cloud server 104 based on the automatic machine learning algorithm.
[0115] It can be understood that in the power prediction model set, each new energy station corresponds to at least one optional power prediction model. The optional power prediction model can be in the form of a physical model, a neural network model, a deep learning model, etc., and the optional power prediction models corresponding to the new energy station can be constructed based on different algorithms. Accordingly, different algorithms can correspond to different training samples.
[0116] In one possible implementation, the optional power prediction model is constructed based on multiple groups of {weather forecast data, historical output power} corresponding to the new energy station; in another possible implementation, the optional power prediction model is constructed based on multiple {historical output power} corresponding to the new energy station. The output results of the optional power prediction models constructed by the above two methods are both the predicted output power of the station.
[0117] The amount of historical output power data is used to determine whether the new energy station is a newly built station or a newly expanded station. It is understood that if the amount of historical output power data for the new energy station is small, then this historical output power is insufficient to construct the optional power prediction model corresponding to the new energy station. In this case, when selecting the target power prediction model for the new energy station, the transfer learning mode is entered; in transfer learning mode, the target power prediction model for the target new energy station can be selected from the optional prediction power models corresponding to other new energy stations.
[0118] The missing future numerical weather forecast data is used to determine whether the currently acquired future numerical weather forecast data can support the optional power prediction model to perform normal forecasting tasks. It is understood that if it cannot support it, the system enters the intelligent backup mode when selecting the target power prediction model. In this intelligent backup mode, based on the missing future numerical weather forecast data, the target power prediction model corresponding to the target new energy station that can cope with the corresponding data missing situation is selected from the various optional prediction power models corresponding to the target new energy station.
[0119] For example, the optional prediction power models corresponding to the target new energy site, after being divided into types and levels, can include at least the following five types: optimal model A1 (input includes weather forecast data and historical output power data), suboptimal model B1 (input includes weather forecast data and historical output power data), optimal model A2 (input includes only weather forecast data, excluding historical output power), suboptimal model B2 (input includes only weather forecast data, excluding historical output power), and no NWP model (input includes only historical output power data, excluding weather forecast data). Among them, suboptimal model B1 and optimal model A1 can be constructed based on different algorithms, and optimal model A2 and suboptimal model B2 can also be constructed based on different algorithms. Specifically, the "smart spare tire mode" provides multiple backup prediction mechanisms, using the lack of future numerical weather forecast data as a selection basis to determine whether to select the optimal model A1, suboptimal model B1, optimal model A2, suboptimal model B2, or no NWP model as the target power prediction model corresponding to the target new energy site. The purpose is to be able to perform power prediction even when there is a disruption in historical output power or when there is no NWP.
[0120] Furthermore, if there is a lack of future numerical weather forecast data and the amount of historical output power data indicates that the target station is not a newly built station or a non-expanded station, then the "normal prediction mode" is entered; in the "normal prediction mode", the target power prediction model corresponding to the target new energy station is selected from the optional prediction power models corresponding to the target new energy station, for example, the optimal model A1 and the suboptimal model B1 in the above example.
[0121] Optionally, each optional power prediction model in the power prediction model set may be pre-delivered to the edge server 102, or may be delivered to the edge server 102 in response to a power prediction requirement of a target new energy site. Subsequently, when performing power prediction, each new energy site in the site set selects its corresponding target power prediction model from the power prediction model set in the edge server 102, achieving precise matching between "sites and models" and providing a more accurate prediction model for the power prediction of the new energy sites.
[0122] S203: Determine a target operating mode of a target power prediction model according to power prediction requirements.
[0123] The target working mode includes the forecast business object, forecast type, and forecast time scale. The forecast business object refers to the type of station performing the forecast task, whether it is a new energy substation, new energy main station, or distributed new energy. Distributed new energy refers to small wind power and photovoltaic power sources distributed at or near the user site. The forecast type refers to the type of forecast business object, such as wind power forecast and photovoltaic power forecast. The forecast time scale includes short-term, ultra-short-term, small time scale power, medium-term, and long-term forecast time scales.
[0124] Understandably, traditional technologies typically use a 15-minute prediction timescale for renewable energy. However, from the perspective of reliable power supply for power system operation, if the prediction timescale can be further reduced, for example to 5 minutes, the reliability of the system power supply can be further improved. This embodiment proposes to reduce the ultra-short-term power prediction timescale from 15 minutes to a small 5-minute timescale based on massive spatiotemporal data and meteorological feature tracking, utilizing ensemble learning and cloud tracking technology, thereby achieving ultra-short-term refined power forecasting at the 5-minute level.
[0125] Furthermore, the target working mode may also include the model output result type. For example, the output result type may be point prediction, interval prediction and probability prediction; wherein, interval prediction and probability prediction are based on the results of point prediction, and interval prediction results and probability prediction results are generated through interval estimation and probability estimation methods.
[0126] Specifically, after selecting the target power prediction model corresponding to the target new energy site from the power prediction model set based on the missing future numerical weather forecast data and the amount of historical output power data, the power prediction demand is analyzed and the target working mode of the target power prediction model is selected from the candidate working modes obtained by combining the prediction business object, prediction type and prediction time scale.
[0127] Each station is distributed in a different location and therefore corresponds to a different scenario. New energy stations in different scenarios may correspond to different prediction models in the power prediction model set (prediction model library). As shown in FIG3 , during the prediction model design phase, new energy stations in different scenarios are designed based on a unified “prediction model design process” to obtain their respective corresponding models. The models developed and designed in the “prediction model design process” may be physical models, statistical models, or artificial intelligence models. During the design process, the adaptability of the above-mentioned different types of models (i.e., physical models, statistical models, or artificial intelligence models) to different scenarios is determined (i.e., the prediction accuracy reaches the corresponding accuracy threshold), and multiple iterative prediction models corresponding to each scenario are obtained. Then, a prediction model library is constructed based on the multiple iterative prediction models corresponding to each scenario. For example, each iterative prediction model in the prediction model library may be constructed based on a long short-term memory network (LSTM), BP (Back Propagation), random forest, resource matching algorithm, or the like. It is understandable that when determining the prediction model corresponding to any station, the iterative prediction model with the highest prediction accuracy for that station may be selected from the prediction model library.
[0128] S204, according to the target working mode, the target power prediction model is adjusted, and based on the adjusted target power prediction model, the target output power of the target new energy station in the future period is predicted according to the future numerical weather forecast data of the future period and the historical output power of the historical period corresponding to the future period.
[0129] Among them, after adjusting the working mode of the target power prediction model to the target working mode, the future numerical weather forecast data and historical output power are input into the target power prediction model to obtain the target output power of the future period output by the target power prediction model.
[0130] Exemplarily, if the target power prediction model is the above-mentioned optimal model A1 or suboptimal model B1, then {future numerical weather forecast data, historical output power} is input into the target power prediction model; if the target power prediction model is the above-mentioned optimal model A2 or suboptimal model B2, then {future numerical weather forecast data is input into the target power prediction model; if the target power prediction model is the above-mentioned non-NWP model, then {historical output power} is input into the target power prediction model, and the corresponding target output power can be obtained.
[0131] In the above-mentioned automatic machine learning cloud-edge collaborative prediction method for new energy power, in the new energy system, the usage status of each station (reflected by the amount of historical output power data) is different, and the same station uses different future numerical weather forecast data for power prediction in different time periods. The unified power prediction model used in traditional technology cannot adapt to different stations; in this application, based on the station usage status and future numerical weather forecast data missing conditions corresponding to the target new energy station, the target power prediction model corresponding to the target new energy station is accurately matched from the power prediction model set. The target power prediction model can perform power prediction when adapted to the corresponding weather forecast data missing conditions or station usage status, so that accurate prediction of the output power of different stations can be achieved; further, based on the power prediction requirements of the target new energy station, the working mode of the target power prediction model is adjusted to obtain a target output power that is more in line with the prediction requirements, thereby improving the accuracy of power prediction.
[0132] In one embodiment, this embodiment provides an optional method for adjusting the target power prediction model according to the target operating mode, and based on the adjusted target power prediction model, predicting the target output power of the target new energy station in the future period according to the future numerical weather forecast data of the future period and the historical output power of the historical period corresponding to the future period, that is, providing a method for refining S104. The specific implementation process may include: if the station type of the target new energy station is distributed new energy, then obtaining the grid location of the target new energy station; based on the adjusted target power prediction model, according to the historical output power and the future numerical weather forecast data corresponding to the grid location, predicting the target output power of the target new energy station in the future period.
[0133] Among them, the distributed new energy in the new energy system is arranged in a mesh structure, and each distributed new energy corresponds to a grid position. Accordingly, each grid position corresponds to corresponding future numerical weather forecast data.
[0134] In this embodiment, when the target new energy station is a distributed new energy station, differentiated and refined power prediction can be achieved by matching the future numerical weather forecast data of the grid location where the distributed new energy station is located.
[0135] As shown in FIG4 , this embodiment provides an optional method for selecting a target power prediction model corresponding to a target new energy station from a power prediction model set based on the lack of future numerical weather forecast data and the amount of historical output power data, that is, a method for refining S102. The specific implementation process may include:
[0136] S401, when it is identified that the target new energy station is a newly built station or a newly expanded station based on the amount of historical output power data, the data missing level of future numerical weather forecast data is determined based on the missing status of future numerical weather forecast data.
[0137] The data missing level may include a sporadic missing level and a continuous missing level; optionally, the sporadic missing level corresponds to a short interruption (e.g., less than or equal to 2 hours), and the continuous missing level may correspond to a continuous missing of NWP data for more than 7 days.
[0138] Optionally, if the amount of historical output power data reaches the construction quantity threshold, the target new energy station is determined to be a non-newly built station or a non-expanded station. It can be understood that the non-newly built station or the non-expanded station has a corresponding constructed optional power prediction model in the power prediction model set; if the amount of historical output power data corresponding to the target new energy station does not reach the construction quantity threshold, the target new energy station is determined to be a new station or an expanded station. It can be understood that the new station or the expanded station does not have a corresponding optional power prediction model in the power prediction model set, and the "transfer learning mode" is entered at this time; in the "transfer learning mode", the optional power prediction model corresponding to other new energy stations (non-newly built stations or non-expanded stations) is used as the target power prediction model of the target new energy station to implement the "migration" action; among them, each other new energy station is a station in the station set except the target new energy station.
[0139] S402 : Based on the basic information of the target new energy station, a similar station whose similarity to the target new energy station reaches a similarity threshold is selected from other new energy stations.
[0140] Optionally, when determining similar sites (i.e., the refinement process of S402), as shown in FIG5 , the following process may be specifically included:
[0141] S501, obtaining basic information of the target new energy station and basic information of other new energy stations.
[0142] The basic information includes at least one dimension of data among longitude and latitude data, capacity data, terrain data, landform data and altitude data.
[0143] S1002 , determining a similarity value between the target new energy station and each other new energy station in each dimension of data based on the basic information of the target new energy station and the basic information of each other new energy station.
[0144] Optionally, when determining the similarity value between the target new energy station and any other new energy station in each dimension of data, the following process is specifically included:
[0145] Sa. Calculate similarity distance: Use the following formulas (1)-(3) to standardize the numerical features in the above basic information: f imax =max(f i ) (2) f imin =min(f i ) (3)
[0146] Among them, f i is the numerical characteristic of the i-th station (the i-th station can be the target new energy station or other new energy stations); f imax is all f i The maximum value in f imin is all f i The minimum value in .
[0147] Sb, using the following formula (4), calculates the similarity value between the target new energy station and any other new energy station in each dimension of data:
[0148] Among them, the weight coefficient w i ≥0, p is the norm identifier, x,y i are the basic information matrices of the station to be transferred for learning (target new energy station) and other new energy stations, x n represents the nth feature of the target new energy station, y jn represents the nth feature of the jth other new energy station in the station cluster; dist(x,y i ) is the similarity value of the target new energy station and any other new energy station in each dimension of data.
[0149] S1003 , based on the similarity values between the target new energy station and other new energy stations in each dimension of data, selecting a similar station from other new energy stations whose similarity with the target new energy station reaches a similarity threshold.
[0150] Specifically, calculate the similarity value dist(x,y i ), dist(x,y i ) are considered similar stations, with at least one similar station. For example, in this embodiment, three similar stations are selected, and the optional power prediction models corresponding to these three similar stations are formed into a migration model library. This migration model library is then "migrated" to the target new energy station. Because the data types and formats of future numerical weather forecast data and historical output power corresponding to new energy stations are unified, the migration model library is highly compatible and adaptable to all stations.
[0151] S403 , selecting a target power prediction model corresponding to the target new energy station from the optional power prediction models belonging to similar stations in the power prediction model set according to the data missing level.
[0152] Specifically, the above-mentioned "migration model library" is used to predict the target output power of the target new energy station. When predicting the target output power, the corresponding optimal model A1, suboptimal model B1, optimal model A2, suboptimal model B2 or no NWP model is selected from the "migration model library" as the target power prediction model of the target new energy station according to the data missing level.
[0153] As shown in FIG6 , this embodiment provides an optional method for selecting a target power prediction model corresponding to a target new energy station from a power prediction model set based on the lack of future numerical weather forecast data and the amount of historical output power data, that is, a method for refining S102. The specific implementation process may include:
[0154] S601, when it is identified that the target new energy station is not a newly built station or a non-expanded station based on the amount of historical output power data, the data missing level of future numerical weather forecast data is determined based on the missing status of future numerical weather forecast data.
[0155] It can be understood that if the target new energy station is not a newly built station or a non-expanded station, then the non-newly built station or the non-expanded station has a corresponding optional power prediction model in the power prediction model set, and each optional power prediction model is, for example, the above-mentioned optimal model A1, suboptimal model B1, optimal model A2, suboptimal model B2 and no NWP model.
[0156] S602 : Selecting a target power prediction model from the optional power prediction models belonging to the target new energy station in the power prediction model set according to the data missing level.
[0157] In one implementation, if the data missing level is a fragmented missing level, the missing data in the future numerical weather forecast data is supplemented, and based on the adjusted target power prediction model, the target output power of the target new energy station in the future period is predicted based on the historical output power and the supplemented future numerical weather forecast data. Specifically, if it is detected that the data missing period of the future numerical weather forecast data is less than a preset interruption period (e.g., 2 hours), the data missing level is determined to be a fragmented missing level.
[0158] When filling in missing data in future numerical weather forecast data, a neighbor filling method or an intelligent filling method can be used. The neighbor filling method includes filling in missing data in future numerical weather forecast data based on candidate numerical weather forecast data corresponding to each historical moment within a preset time range. As shown in FIG7 , exemplarily, the historical moments within the preset time range are t1 and t2, t1 corresponds to candidate numerical weather forecast data 1, t2 corresponds to candidate numerical weather forecast data 2, and t3 is the future numerical weather forecast data. When t3 is the future numerical weather forecast data and there are missing values 1 and 2 in the missing value 1, the missing moment 1 corresponding to the missing value 1 and the missing moment 2 corresponding to the missing value 2 are determined. Then, first, it is queried in the candidate numerical weather forecast data 1 whether the data corresponding to the missing moment 1 is missing. As shown in FIG7 , the data corresponding to the missing moment 1 in the candidate numerical weather forecast data 1 is missing value 1', then it is queried in the candidate numerical weather forecast data 2 whether the data corresponding to the missing moment 1 is missing. As shown in FIG7 , the data corresponding to the missing moment 1 in the candidate numerical weather forecast data 2 is not missing data, then the non-missing data is supplemented to the future numerical weather forecast data. The processing method for the missing value 2 is the same as the processing method for the missing value 1. Finally, the supplemented future numerical weather forecast data, that is, the future numerical weather forecast data corresponding to the moment tx, is obtained.
[0159] It can be understood that if the data missing level is a scattered missing level, then the target power prediction model with input data of {historical output power, weather forecast data} is selected from the optional power prediction models belonging to the target new energy station in the power prediction model set, such as the above-mentioned optimal model A1 and suboptimal model B1. Based on the completed future numerical weather forecast data, the target output power of the target new energy station in the future period can be predicted.
[0160] In this embodiment, when there is a temporary interruption or omission in the weather forecast data, the weather forecast data is automatically supplemented, and a target power prediction model whose input data is {historical output power, weather forecast data} is automatically selected, thereby realizing automatic processing of temporary data omissions.
[0161] In another possible implementation, if the data missing level is a continuous missing level, the target output power of the target new energy station in the future period is predicted based on the adjusted target power prediction model and the historical output power.
[0162] Specifically, if it is detected that the data missing period of future numerical weather forecast data reaches more than 7 days, the data missing level is determined to be a continuous missing level, and the prediction model will enter the "smart spare tire mode" at this time.
[0163] It is understood that if the data missing level is a continuous missing level, then the target power prediction model whose input data is {historical output power} is selected from the optional power prediction models belonging to the target new energy station in the power prediction model set, such as the above-mentioned non-NWP model. As shown in Figure 8, this non-NWP model can be constructed based on the power shift algorithm, the time series decomposition method, and the genetic optimization algorithm; wherein, the principle of the power shift algorithm is to replace the predicted power with the adjacent historical power; it is understood that the parameters refer to the unknowns in the algorithm, some of which need to be specified manually, such as the learning efficiency in the neural network algorithm, and some are fitted from the data, such as the coefficients in linear regression, and so on. When using the selected algorithm for modeling, the set or obtained parameters are likely not optimal or near-optimal. In this case, the parameters need to be optimized to obtain a better prediction model. The essence of the genetic optimization algorithm is to select a batch of optimal parameters to optimize the objective function. Optionally, in this embodiment, an optimization model library is first constructed based on various different algorithms, and then the optimal model is selected from the optimization models according to the genetic optimization algorithm, wherein the models in the optimization model library can be: generalized linear regression model (GLS), generalized linear model (GLM), VAR model (Vector autoregressive model), etc.; the objective function in this embodiment is a function that makes the predicted output power of the model closer to the actual output power. Exemplarily, the objective function can be: the root mean square error relative to the startup capacity, the prediction qualification rate, the root mean square error relative to the actual power, the maximum absolute prediction deviation during the peak load period, and the maximum absolute prediction deviation of the day, etc.
[0164] Specifically, the input data for prediction using the aforementioned non-NWP model is all {historical output power}. Therefore, based on the historical output power, the target output power for the target new energy station in the future period can be predicted. Furthermore, as shown in Figure 9, in "Smart Spare Tire Mode," the "Multiple Alternative Mechanism" is enabled to autonomously call multiple priority prediction models, namely, the "optimal model" (input includes historical output power), the optimal model (input excludes historical output power), the "suboptimal model" (input includes historical output power), the "suboptimal model" (input excludes historical output power), and the "non-NWP model."
[0165] It is understandable that in the "smart spare tire mode", there are three calling situations:
[0166] (1) If there is no missing NWP data, the order of calling the optimal model and the suboptimal model is determined according to the interruption of historical output power. For example, in Figure 9, if there is no interruption in historical output power, the "optimal model A1" (input includes historical output power) or the "suboptimal model B1" (input includes historical output power) is called first; if there is an interruption in historical output power, the "optimal model A2" (input does not include historical output power) or the "suboptimal model B2" (input does not include historical output power) is called;
[0167] (2) NWP data are missing, and NWP data are available after the missing data are filled: In this case, the missing NWP data can be filled by the above-mentioned short-term interruption intelligent filling method or the continuous interruption intelligent filling method. After filling, it is considered that the NWP data are not missing, and the calling order of the optimal model and the suboptimal model is the same as in the above case (1);
[0168] (3) If NWP data are missing and the NWP data are still unavailable after filling the missing data: If it is determined that there is no interruption in historical output power, the non-NWP model is called.
[0169] In this embodiment, when there is no missing NWP data (or when there is a short interruption), the model with historical output power or without historical output power can be preferentially called according to the intelligent spare tire mode; when there is a continuous lack of NWP data, the target power prediction model with the input data of {historical output power} is automatically selected, realizing automatic processing of the continuous missing situation.
[0170] In one embodiment, as shown in FIG10 , a cloud-edge collaborative prediction method for renewable energy power based on automatic machine learning is provided. The method is described by applying it to the cloud server 104 in FIG1 as an example, including the following steps:
[0171] S1001, for each new energy station in the station set, train each initial model based on the training sample set corresponding to the new energy station, and select an optional power prediction model corresponding to the new energy station from the trained initial models.
[0172] The training sample set includes multiple sets of {sample weather forecast data, sample output power}, and different initial models are constructed based on different algorithms. For example, the different initial models constructed based on different algorithms can include a variety of prediction models including physical models, statistical models, and more than 10 AI models (as shown in Figure 3).
[0173] Specifically, as shown in FIG11 , each initial model is trained based on the training sample set corresponding to the new energy station, including:
[0174] S1101, performing data mining processing on the original weather data in the original data corresponding to the new energy station to obtain sample weather forecast data in a training sample set.
[0175] The raw data corresponding to each new energy station includes raw weather data (i.e., raw numerical weather forecast data) and raw power data (i.e., historical output power). The amount of raw weather data for each station is relatively large. For example, the southern region is forecast four times a day (forming four numerical weather forecast data), and the amount of data for each forecast for the next seven days is about 100G. In addition, the information in a single raw numerical weather forecast data is complex and the features are redundant, with more than 150 NWP feature quantities. In addition, a large amount of erroneous data, missing data, and power-limited data in the raw power data are difficult to accurately identify, and data noise will seriously reduce the upper limit of prediction accuracy. If all the data and features in the raw weather data are used for algorithm model training and prediction, it will cause problems such as large computational load, high computing power requirements, and lengthy time consumption. It will also interfere with the power prediction model, increase the complexity of model training and prediction, and greatly reduce the accuracy of the algorithm and its generalization ability. However, in the process of fully mining feature information, the "dimensionality curse" is very likely to occur. General data processing methods cannot completely solve the problem of the huge amount of raw numerical data. Therefore, in order to improve the effectiveness of data processing, the embodiment of the present application needs to perform data mining processing on the original data; specifically, as shown in FIG12 , the data mining processing of the original weather data may include:
[0176] S1201, based on the slicing technology, the original weather data in the original data is cleaned.
[0177] This example uses a very fast slicing and cleaning technology based on massive data. This technology can complete the cleaning of 25GB of massive multidimensional meteorological data files with specified coordinates, time periods, and feature quantities in single numerical weather forecast data (original numerical weather forecast data) within 1 hour. The specific steps are as follows:
[0178] Step 01: Extract massive multi-dimensional meteorological data files from the original weather data based on the specified coordinates, time period, feature quantities and other information.
[0179] The specified coordinates are the specified longitude and latitude.
[0180] Step 02: Use slicing technology to slice the multidimensional meteorological data file according to the customized slicing rules to form multiple file slices.
[0181] Step 03: Use a distributed architecture for each file slice to perform data extraction and data mining in parallel to extract the cleaned original weather data.
[0182] The cleaned original weather data includes multiple cleaned file slices; data extraction and data mining can be determined based on a preset data cleaning algorithm, with the goal of ensuring that the cleaned file slices contain corresponding target data information.
[0183] Step 04: Load the target data information in each cleaned file slice from the memory, splice and output it to a unified table and database for storage.
[0184] Then, the target data information in each file slice after cleaning is further extracted to obtain sample weather forecast data. Specifically, the further feature extraction is a meteorological feature cross-recursive optimization step. The implementation process of this step is: through cross-validation and recursive elimination of algorithms such as correlation coefficients and random forests, more than 150 complex NWP data items in the target data information are screened for feature quantities, and 4 to 5 key feature quantities that have a prominent effect on accuracy improvement are located, which reduces the complexity and computational complexity of the power prediction model, reduces the interference of irrelevant meteorological features on prediction accuracy, and improves the accuracy of new energy power prediction. Specifically, it can include:
[0185] S1202: Perform correlation analysis on the original output power and the cleaned original weather data to determine a principal component feature set in the original weather data.
[0186] Specifically, the specific implementation process of the above S1202 is as follows:
[0187] Step 11: Based on the machine learning algorithm, all features in the cleaned raw weather data are preliminarily screened, features below the score screening threshold are discarded, and the remaining features are determined as the first feature set.
[0188] The machine learning algorithm can be the Random Forest algorithm, a supervised learning algorithm based on decision trees. It is an ensemble learning algorithm that classifies and filters all features based on their importance. The threshold, which serves as a feature engineering input parameter and can be manually adjusted, defaults to 0.05.
[0189] Step 12: Based on correlation analysis, calculate the correlation coefficients between the features of all dimensions in the filtered original weather data and the original output power, discard the features whose absolute values of the correlation coefficients are lower than the correlation threshold, and determine the remaining features as the second feature set.
[0190] The correlation coefficient may be a phase correlation coefficient obtained by taking its absolute value.
[0191] Specifically, when performing correlation analysis, you can call the corr() function in Python to form a correlation analysis matrix, and then perform subsequent analysis. This is equivalent to directly inputting a data matrix and outputting a matrix of correlation values.
[0192] Optionally, the correlation threshold is used as an input parameter for feature engineering and can be manually adjusted. The default values are 0.2 for photovoltaic power and 0.45 for wind power.
[0193] Step 13: Based on the union of the first feature set and the second feature set, a principal component feature set in the original weather data is formed.
[0194] S1203, verify the validity of the principal component feature set to obtain a valid feature set.
[0195] Specifically, K-fold verification and recursive elimination method are used to verify the validity of the principal component feature set to obtain a valid feature set.
[0196] S1204: Use the valid feature set as sample weather forecast data in the training sample set.
[0197] Finally, the valid feature set is used as the return value of feature engineering in the form of a list and as the sample weather forecast data in the training sample set for training the power model.
[0198] After S1101, the sample weather forecast data in the training sample set is obtained. Then, the original data is processed to obtain the sample output power in the training sample set. Specifically, the following process is included:
[0199] S1102 , performing multi-stage noise reduction processing on the original output power in the original data to obtain the sample output power in the training sample set.
[0200] Optionally, the step of performing multi-stage noise reduction on the original output power in S1102 specifically includes the following process:
[0201] Step 21: Identify and correct obvious abnormal data in the original power data, such as data that exceeds the installed capacity of new energy stations and the original output power is negative, and set the power data of photovoltaic power during the night period to zero.
[0202] It is understandable that the negative values are mostly caused by the accuracy of the acquisition equipment. For example, after the hardware ages, its range does not start from zero but from a negative value.
[0203] Step 22: Based on the weather-power correlation calculated between the original weather data and the original power data, a weather-power depth mapping relationship is established.
[0204] Step 23: For any moment or time period, when the original weather data corresponding to the moment or time period is known, the estimated power data corresponding to the moment or time period is estimated based on the weather-power depth mapping relationship.
[0205] Step 24: Compare the estimated power data with the original power data. If the difference between the original power data and the estimated power data is greater than a preset difference, the time or time period is determined to be a power restriction period and a communication interruption period for the new energy station.
[0206] The difference between the original power data and the estimated power data is greater than a preset difference, indicating that the original power data is much lower than the estimated power data.
[0207] Step 25: Correct the original power data during the power-limited period and the communication interruption period to remove data noise and obtain the original power data after multi-stage noise reduction.
[0208] Among them, the original power data after multi-stage noise reduction is the sample output power in the training sample set.
[0209] Specifically, the process of the data cleaning step, the multi-stage noise reduction step, and the meteorological cross-recursive optimization step based on the slicing technology is shown in FIG13 .
[0210] S1103: Training each initial model based on the training sample set.
[0211] Specifically, the training process in this embodiment is automatically executed by a full-cycle automatic machine learning algorithm. The training link in the full-cycle automatic machine learning algorithm can automatically perform variable step-size parameter optimization, automatically adjust the machine learning model parameters, and carry out multi-model ensemble learning based on the optimal parameters, and automatically train and match the optimal model suitable for different prediction business objects. As shown in Figure 14, it can specifically include the following processes:
[0212] S1401: For any initial model, obtain the initial hyperparameter space and initial step size corresponding to the initial model.
[0213] S1402: Adjust the initial step size based on the number of iterations.
[0214] Specifically, the calculation is based on the number of search iterations, the size of the improvement between each generation, and the data size. If the number of iterations increases and the improvement value decreases, the number of steps decreases.
[0215] In formula (5), k is the number of iterations, f k is the optimization objective function value obtained by the kth iteration. Here, the root mean square error between the predicted result and the actual power is taken as the optimization target; dk is the step size of the kth iteration.
[0216] S1403: Based on the adjusted step size, search in the initial hyperparameter space to obtain a hyperparameter optimization space.
[0217] Among them, in the variable step size optimization of model hyperparameters, a large step size is first used to conduct a preliminary search for the reasonable selection range of the initial model hyperparameters (hyperparameter initial space) to reduce the parameter search space and obtain the hyperparameter optimization space.
[0218] Among them, the maximum step size can be set according to experience.
[0219] S1404: In the hyperparameter optimization space, based on the initial hyperparameter value and the Bayesian optimization algorithm, obtain at least one optimized hyperparameter to be verified.
[0220] Specifically, within the hyperparameter optimization space, a small step size is used to perform grid optimization on the hyperparameters to find the hyperparameters that make the initial model perform best on the experimental dataset, and an automatic optimization mechanism for the model hyperparameters is established to improve the prediction accuracy.
[0221] As shown in Figure 15, the above-mentioned model hyperparameter optimization process (S1401 random optimization, S1402 determination of optimization scale, S1403 narrowing of optimization range, S1404 grid optimization) obtains the optimized model parameters. Based on the optimized model parameters and various algorithms, a physical model, an artificial intelligence model and a deep learning model are constructed, and then the physical model, the artificial intelligence model and the deep learning model are integrated and learned. Integrated learning refers to a decision (power prediction) made jointly by multiple decision makers (multiple models).
[0222] Optionally, this embodiment uses a sequential model optimization algorithm within the hyperparameter optimization space to solve the non-differentiable, discontinuous, and high-cost objective function of predictive model hyperparameter optimization. In the Bayesian optimization of hyperparameters, the basic steps are as follows:
[0223] Step 31. Initialize the error probability distribution p(y|x) of model M under a set of hyperparameters x, where y is the actual value of the objective function using the hyperparameter set x.
[0224] Step 32: Construct a new set of hyperparameters x*, as shown in the following formula (6): x*=argminS(x,M t -1,H) (6)
[0225] Where S is the acquisition function.
[0226] Step 33. Calculate the objective function f(x*) under this hyperparameter set.
[0227] Step 34. Update the sample set \(H = H\cup(x^*, f(x^*))\).
[0228] Step 35. Estimate the new \(p(y|x, H)\) based on \(H\).
[0229] Step 36. Loop from Step 32 until the maximum number of iterations is reached.
[0230] It can be understood that Bayesian optimization uses the previously calculated points \(x\), calculates the posterior expectation value of the objective function \(f\), and uses a specific sampling function to calculate the objective function \(f\) at the new \(x^*\) to minimize the newly calculated \(f\). Repeat the above steps (Step 31 - Step 36) until the maximum number of iterations is reached or the convergence condition is met.
[0231] The estimation of the prior probability \(p(y|M)\) generally has the Gaussian method and the tree - structured Parzen estimator method (TPE). In this embodiment, the latter is adopted.
[0232] The TPE algorithm defines \(y^*\) as the quantile \(\gamma\) of the observed \(y\) values, so \(\gamma = p(y < y^*)\). To find the optimal sampling point of the objective function, the most commonly used is the Expected Improvement (EI), and the formula is as follows - Equation (7):
[0233] where \(l(x)\) is the probability density function formed by using the observed values \(\{x^{(i)}\}\) such that the corresponding loss \(f(x^{(i)})\) is less than \(y^*\), and \(g(x)\) is the probability density function formed by using the remaining observed values.
[0234] As can be seen from the above formula, to maximize EI, \(l(x)\) should be maximized and \(g(x)\) should be minimized. In the SMBO (Sequential Model - Based Optimization for General Algorithm Configuration) algorithm, TPE can find the next set of hyperparameters \(x^*\) corresponding to maximizing EI. Finally, the hyperparameter set \(x^*\) corresponding to minimizing the objective function \(f(x)\) is obtained through iteration.
[0235] S1405. Based on the training sample set corresponding to the new - energy power station and each hyper - optimized parameter to be verified, perform k - fold cross - validation on the initial model to complete the training of the initial model.
[0236] Optionally, in this embodiment, during the process of Bayesian hyperparameter optimization, the training iteration of the model is carried out in combination with k - fold cross - validation.
[0237] Specifically, k-fold cross-validation means that during the initial sampling, the training sample set is divided into k sub-samples. A single sub-sample is retained as data for the validation model, and the other k-1 samples are used to train the model. Cross-validation is repeated k times, with each sub-sample validated once, and the model with the smallest root mean square error is taken as the final model. The advantage of this method is that by averaging the results of k different group training to reduce variance, the performance of the model is less sensitive to the division of the data, which can avoid model overfitting and enhance the generalization ability of the model. The steps are as follows:
[0238] In the first step, the training sample set is randomly divided into k parts without repeated sampling.
[0239] In the second step, one of the samples is selected as the validation set each time, and the remaining k-1 samples are used as the training set for training the initial model.
[0240] In the third step, repeat the second step k times, so that each subset has one chance to be used as the validation set and the rest as the training set. After training on each training set, a model is obtained. This model is used to make power predictions on the corresponding validation set, and the model evaluation metrics are calculated and saved.
[0241] The fourth step is to rank the k groups of test results and take the model with the smallest root mean square error as the training model output under k-fold cross-validation, and as the optimal model evaluated during Bayesian hyperparameter optimization in the current number of iterations.
[0242] It can be understood that the above-mentioned optimal model can be the optimal model A1 in the above example; the suboptimal model B1 corresponding to the suboptimal hyperparameters and the optimal NWP-free model can also be obtained through the above method.
[0243] S1002: Construct a power prediction model set based on the optional power prediction models corresponding to each new energy station.
[0244] After S1001 completes the training of each initial model, the model with the highest prediction accuracy for each new energy station is adaptively selected based on different station types, topography, and climate conditions. Taking any new energy station in S1001 as an example, as shown in FIG16 , the specific process may include the following:
[0245] S1601: Input the verification sample set corresponding to the new energy station into each trained initial model to obtain the verification output power output by each trained initial model.
[0246] S1602: Determine the prediction accuracy of each trained initial model based on the actual output power corresponding to the verification sample in the verification sample set and the verification output power output by each trained initial model.
[0247] S1603: Select an optional power prediction model corresponding to the new energy station from the trained initial models based on the prediction accuracy.
[0248] In this embodiment, an initial model with a prediction accuracy that meets a preset precision is specifically matched for each new energy station as an optional power prediction model for the new energy station, thereby realizing the function of building a high-precision optional power prediction model for each new energy station.
[0249] Furthermore, as time goes by, the practicality of the original model of some new energy stations decreases due to adjustments in operating modes or expansion and transformation. On the other hand, as new energy stations accumulate more historical operating data, they provide a richer sample set for model optimization training. Therefore, in order to continuously improve the accuracy of new energy power prediction at the station, it is necessary to regularly backtrack the optional power prediction model corresponding to each new energy station, continuously match the trained initial model corresponding to each new energy station in the background, calculate the prediction accuracy of the initial model, update the current optimal prediction model, and train and match different prediction models for each new energy station in different time periods to achieve dynamic adaptation of the model.
[0250] In this embodiment, the short-term forecast model can be updated on a calendar year basis, and the ultra-short-term forecast model can be updated on a quarterly basis. This allows each new energy station to consistently customize the model with the best forecast accuracy, achieving differentiated forecasting. As shown in Figure 17, at time t1, each station has its own optimal model. These models are continuously and dynamically updated. For example, if the update interval is (tn-t1), the updated optimal model for each station can be obtained at time tn.
[0251] After the power prediction model set is obtained through training in S1001 and S1002, it also includes:
[0252] S1003: Send the power prediction model set to the edge server.
[0253] Among them, the cloud-edge collaborative prediction system in this embodiment is divided into three different versions of the system according to the business: one is the cloud version of the prediction system, which is deployed in the power grid cloud IV area and is responsible for the construction of the full-cycle automatic machine learning sample library and the full-cycle automatic machine learning training; the second is the master station version of the prediction system, which is deployed in the power dispatching master station, power generation group prediction center, etc., and is responsible for the full-cycle automatic machine learning prediction at the master station end; the third is the substation version of the prediction system, which is deployed at the new energy station and directly undertakes the power prediction business of the new energy station.
[0254] Optionally, the automatic machine learning new energy power cloud-edge collaborative prediction method in this embodiment is implemented based on the full-cycle machine learning technology algorithm and based on the full-cycle machine learning technology architecture.
[0255] As shown in Figure 18, the full-cycle machine learning technology algorithm architecture includes three parts: full-cycle automatic machine learning sample library construction, full-cycle automatic machine learning training, and full-cycle automatic machine learning prediction. Among them, the sub-steps corresponding to the ultra-fast slice cleaning in the construction of the full-cycle automatic machine learning sample library specifically refer to the above S1201, the sub-steps corresponding to the multi-stage noise reduction in the construction of the full-cycle automatic machine learning sample library specifically refer to the above S1102, and the sub-steps corresponding to the adaptive selection of NWP data features in the construction of the full-cycle automatic machine learning sample library specifically refer to the above S1202, S1203 and S1204; further, the sub-steps corresponding to the variable step size optimization of hyperparameters in the full-cycle automatic machine learning training specifically refer to the above S1401 to S1405, and the sub-steps corresponding to the multi-model ensemble learning in the full-cycle automatic machine learning training specifically refer to the above S1002; further, the sub-steps corresponding to the transfer learning in the full-cycle automatic machine learning prediction specifically refer to the above S501 to S503; the sub-steps corresponding to the normal prediction mode in the full-cycle automatic machine learning prediction specifically refer to the above S204; the sub-steps corresponding to the smart spare tire in the full-cycle automatic machine learning prediction specifically refer to the process shown in Figure 9 above.
[0256] Specifically, as shown in Figure 19, when performing target power forecasting, the new energy forecasting system (cloud version) is deployed in the grid cloud security zone IV, pulling NWP data (future numerical weather forecast data) from the grid data center. After training in the cloud, the model and the NWP data of specific sites are pushed to the internal and external network exchange platform (unstructured channel). The security zone V pulls the target forecasting model and NWP data via the unstructured channel and pushes them to the grid public cloud. As a result, master stations and sites for specific forecasting services can pull the NWP data and target forecasting models required for specific services from the public cloud, and carry out forecasting services at the master stations, sites, and other business terminals, greatly leveraging the advantages of cloud-edge collaborative new energy power forecasting.
[0257] Figure 20 shows a schematic diagram of the edge-cloud collaboration system framework. This system framework includes a data source input layer, a basic data processing layer (Hive, Spark SQL), a core business layer, a result output layer, and downstream systems. Furthermore, it also provides a monitoring system, data monitoring, and alarm services.
[0258] Among them, the data source input layer includes: meteorological data center module, text upload module, dispatch center module, new energy station module, and other data source modules; the basic data processing layer (Hive, Spark SQL) includes data cleaning module, data splitting module, data integration module and feature statistics module; the core business layer includes station power prediction module, station model training module, transfer learning module and timing scheduling module; the result output layer includes synchronization data warehouse module, synchronization MySQL module, synchronization Hbase module, STEP service module and HTTP interface module; downstream systems include power prediction Web system, master station power prediction system, substation power prediction system, and other systems.
[0259] It should be understood that, although the steps in the flowcharts of the above embodiments are shown in sequence as indicated by the arrows, these steps are not necessarily performed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be performed in other orders. Moreover, at least a portion of the steps in the flowcharts of the above embodiments may include multiple steps or multiple stages, and these steps or stages are not necessarily performed at the same time, but can be performed at different times. The execution order of these steps or stages is not necessarily to be performed in sequence, but can be performed in turn or alternately with other steps or at least a portion of steps or stages in other steps.
[0260] In one embodiment, a new energy power cloud-edge collaborative prediction system based on automatic machine learning is provided, comprising: a cloud server and an edge server;
[0261] The cloud server is configured to train each initial model for each new energy station in the station cluster based on a training sample set corresponding to the new energy station, and select an optional power prediction model corresponding to the new energy station from the trained initial models; wherein the training sample set includes sample weather forecast data and sample output power, and different initial models are constructed based on different algorithms; construct a power prediction model set based on the optional power prediction models corresponding to each new energy station; and distribute the power prediction model set to the edge server;
[0262] The edge server is used to receive a power prediction model set; in response to the power prediction demand for the target new energy station, obtain the future numerical weather forecast data of the target new energy station in the future period and the historical output power of the historical period corresponding to the future period; according to the missing situation of future numerical weather forecast data and the amount of historical output power data, select the target power prediction model corresponding to the target new energy station from the power prediction model set; wherein, each optional power prediction model in the power prediction model set is trained and issued by the cloud server based on the automatic machine learning algorithm; according to the power prediction demand, determine the target working mode of the target power prediction model; wherein, the target working mode includes the prediction business object, the prediction type and the prediction time scale; according to the target working mode, adjust the target power prediction model, and based on the adjusted target power prediction model, predict the target output power of the target new energy station in the future period according to the future numerical weather forecast data in the future period and the historical output power of the historical period corresponding to the future period.
[0263] Based on the same inventive concept, the embodiment of the present application also provides an automatic machine learning new energy power cloud-edge collaborative prediction device 1 for implementing the automatic machine learning new energy power cloud-edge collaborative prediction method involved above. The implementation solution provided by the device is similar to the implementation solution recorded in the above method, so the specific limitations in the embodiments of one or more automatic machine learning new energy power cloud-edge collaborative prediction devices provided below can be found in the above limitations on an automatic machine learning new energy power cloud-edge collaborative prediction method, which will not be repeated here.
[0264] In one embodiment, as shown in FIG21 , a cloud-edge collaborative prediction device 1 for new energy power based on automatic machine learning is provided, comprising: a response module 11, a selection module 12, an adjustment module 13, and a prediction module 14, wherein:
[0265] A response module 11 is configured to respond to a power forecast demand for a target new energy station and obtain future numerical weather forecast data of the target new energy station in a future period and historical output power of a historical period corresponding to the future period;
[0266] A selection module 12 is configured to select a target power prediction model corresponding to a target new energy station from a power prediction model set based on the lack of future numerical weather forecast data and the amount of historical output power data; wherein each selectable power prediction model in the power prediction model set is trained and issued by a cloud server based on an automatic machine learning algorithm;
[0267] The adjustment module 13 is used to determine the target working mode of the target power prediction model according to the power prediction requirements; wherein the prediction business object, prediction type and prediction time scale are predicted;
[0268] The prediction module 14 is used to adjust the target power prediction model according to the target working mode, and based on the adjusted target power prediction model, predict the target output power of the target new energy station in the future period according to the future numerical weather forecast data of the future period and the historical output power of the historical period corresponding to the future period.
[0269] In one embodiment, the selection module 12 includes:
[0270] The first screening submodule is configured to determine a data missing level of future numerical weather forecast data based on missing future numerical weather forecast data when the target new energy station is identified as a non-newly built station or a non-expanded station based on the amount of historical output power data;
[0271] The first matching submodule is configured to select a target power prediction model from the optional power prediction models belonging to the target new energy station in the power prediction model set according to the data missing level.
[0272] In one embodiment, the selection module 12 includes:
[0273] The second screening submodule is used to determine the data missing level of future numerical weather forecast data based on the missing status of future numerical weather forecast data when the target new energy station is identified as a newly built station or an expanded station based on the amount of historical output power data;
[0274] A similarity calculation submodule is used to select similar stations from other new energy stations whose similarity with the target new energy station reaches a similarity threshold based on the basic information of the target new energy station; wherein the other new energy stations are stations in the station set other than the target new energy station;
[0275] The second matching submodule is used to select a target power prediction model corresponding to the target new energy site from the optional power prediction models belonging to similar sites in the power prediction model set according to the data missing level.
[0276] In one embodiment, the prediction module 14 includes:
[0277] A completion submodule is used to complete the missing data in the future numerical weather forecast data if the data missing level is a scattered missing level, and based on the adjusted target power prediction model, predict the target output power of the target new energy station in the future period according to the historical output power and the completed future numerical weather forecast data;
[0278] The prediction submodule is used to predict the target output power of the target new energy station in the future period based on the adjusted target power prediction model and the historical output power if the data missing level is a continuous missing level.
[0279] In one embodiment, the similarity calculation submodule is further configured to: obtain basic information of the target new energy station and basic information of each other new energy station; wherein the basic information includes at least one dimension of latitude and longitude data, capacity data, terrain data, landform data, and altitude data;
[0280] Based on the basic information of the target new energy station and the basic information of each other new energy station, determine the similarity value between the target new energy station and each other new energy station in each dimension of data;
[0281] According to the similarity value between the target new energy station and other new energy stations in each dimension of data, similar stations whose similarity with the target new energy station reaches a similarity threshold are selected from other new energy stations.
[0282] In one embodiment, the prediction module 14 is further configured to: if the station type is distributed new energy, obtain the grid location of the target new energy station;
[0283] Based on the adjusted target power prediction model, the target output power of the target new energy station in the future period is predicted according to the historical output power and future numerical weather forecast data corresponding to the grid location.
[0284] In one embodiment, as shown in FIG22 , a new energy power cloud-edge collaborative prediction device 2 based on automatic machine learning is provided, comprising: a training module 21 , a building module 22 , and a sending module 23 , wherein:
[0285] A training module 21 is configured to train each initial model for each new energy station in the station set based on a training sample set corresponding to the new energy station, and select an optional power prediction model corresponding to the new energy station from the trained initial models; wherein the training sample set includes sample weather forecast data and sample output power, and different initial models are constructed based on different algorithms;
[0286] A building module 22 is configured to construct a power prediction model set based on the optional power prediction models corresponding to each new energy station; wherein the power prediction model set is used for the edge server to select a target power prediction model corresponding to the target new energy station based on the lack of future numerical weather forecast data for the target new energy station in the future period and the amount of historical output power data of the target new energy station in the historical period corresponding to the future period; in the target working mode, the target power prediction model is adjusted, and based on the adjusted target power prediction model, the target output power of the target new energy station in the future period is predicted according to the future weather forecast data and the historical output power;
[0287] The sending module 23 is used to send the power prediction model set to the edge server.
[0288] In one embodiment, the training module 21 includes:
[0289] The data mining submodule is used to perform data mining on the original weather data in the original data corresponding to the new energy station to obtain sample weather forecast data in the training sample set;
[0290] The noise reduction processing submodule is used to perform multi-stage noise reduction processing on the original output power in the original data to obtain the sample output power in the training sample set;
[0291] The training submodule is used to train each initial model based on the training sample set.
[0292] In one embodiment, the data mining submodule is further used to: perform data cleaning on the raw weather data in the raw data based on slicing technology;
[0293] Perform correlation analysis on the original output power and the cleaned original weather data to determine the principal component feature set in the original weather data;
[0294] Verify the validity of the principal component feature set to obtain a valid feature set;
[0295] The effective feature set is used as the sample weather forecast data in the training sample set.
[0296] In one embodiment, the training module 21 is further configured to: for any initial model, obtain an initial hyperparameter space and an initial step size corresponding to the initial model;
[0297] Adjusting the initial step size based on the number of iterations;
[0298] Based on the adjusted step size, search the hyperparameter optimization space within the initial hyperparameter space;
[0299] In the hyperparameter optimization space, based on the initial value of the hyperparameter and the Bayesian optimization algorithm, at least one optimized hyperparameter to be verified is obtained;
[0300] Based on the training sample set corresponding to the new energy station and the super-optimization parameters to be verified, the initial model is subjected to k-fold cross-validation to complete the training of the initial model.
[0301] In one embodiment, the training module 21 is further configured to: input the verification sample set corresponding to the new energy station into each trained initial model to obtain the verification output power output by each trained initial model;
[0302] Determine the prediction accuracy of each trained initial model based on the actual output power corresponding to the verification sample in the verification sample set and the verification output power output by each trained initial model;
[0303] According to the prediction accuracy, the optional power prediction model corresponding to the new energy station is selected from the trained initial models.
[0304] Each module in the aforementioned automatic machine learning-based new energy power cloud-edge collaborative prediction device can be implemented in whole or in part through software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in hardware form, or can be stored in a computer device's memory in software form, allowing the processor to call and execute the corresponding operations of each module.
[0305] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as shown in FIG23. The computer device includes a processor, a memory, and a network interface connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store data of a new energy power cloud-edge collaborative prediction method based on automatic machine learning. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it implements a new energy power cloud-edge collaborative prediction method based on automatic machine learning.
[0306] Those skilled in the art will understand that the structure shown in Figure 23 is merely a block diagram of a portion of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.
[0307] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:
[0308] In response to a power forecast demand for a target new energy station, obtaining future numerical weather forecast data of the target new energy station in a future period and historical output power of a historical period corresponding to the future period;
[0309] Based on the lack of future numerical weather forecast data and the amount of historical output power data, a target power prediction model corresponding to the target new energy site is selected from the power prediction model set. Each optional power prediction model in the power prediction model set is trained and issued by the cloud server based on an automated machine learning algorithm.
[0310] Determine the target operating mode of the target power prediction model based on power prediction requirements; wherein the target operating mode includes the prediction business object, prediction type, and prediction time scale;
[0311] According to the target working mode, the target power prediction model is adjusted, and based on the adjusted target power prediction model, the target output power of the target new energy station in the future period is predicted according to the future numerical weather forecast data of the future period and the historical output power of the historical period corresponding to the future period.
[0312] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:
[0313] In response to a power forecast demand for a target new energy station, obtaining future numerical weather forecast data of the target new energy station in a future period and historical output power of a historical period corresponding to the future period;
[0314] Based on the lack of future numerical weather forecast data and the amount of historical output power data, a target power prediction model corresponding to the target new energy site is selected from the power prediction model set. Each optional power prediction model in the power prediction model set is trained and issued by the cloud server based on an automated machine learning algorithm.
[0315] Determine the target operating mode of the target power prediction model based on power prediction requirements; wherein the target operating mode includes the prediction business object, prediction type, and prediction time scale;
[0316] According to the target working mode, the target power prediction model is adjusted, and based on the adjusted target power prediction model, the target output power of the target new energy station in the future period is predicted according to the future numerical weather forecast data of the future period and the historical output power of the historical period corresponding to the future period.
[0317] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:
[0318] For each new energy station in the station cluster, each initial model is trained based on the training sample set corresponding to the new energy station, and an optional power prediction model corresponding to the new energy station is selected from the trained initial models; the training sample set includes sample weather forecast data and sample output power, and different initial models are constructed based on different algorithms;
[0319] A power prediction model set is constructed based on the optional power prediction models corresponding to each new energy site; wherein the power prediction model set is used for the edge server to select a target power prediction model corresponding to the target new energy site based on the lack of future numerical weather forecast data for the target new energy site in the future period and the amount of data on the historical output power of the target new energy site in the historical period corresponding to the future period; in the target working mode, the target power prediction model is adjusted, and based on the adjusted target power prediction model, the target output power of the target new energy site in the future period is predicted according to the future weather forecast data and the historical output power;
[0320] Send the power prediction model set to the edge server.
[0321] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:
[0322] For each new energy station in the station cluster, each initial model is trained based on the training sample set corresponding to the new energy station, and an optional power prediction model corresponding to the new energy station is selected from the trained initial models; the training sample set includes sample weather forecast data and sample output power, and different initial models are constructed based on different algorithms;
[0323] A power prediction model set is constructed based on the optional power prediction models corresponding to each new energy site; wherein the power prediction model set is used for the edge server to select a target power prediction model corresponding to the target new energy site based on the lack of future numerical weather forecast data for the target new energy site in the future period and the amount of data on the historical output power of the target new energy site in the historical period corresponding to the future period; in the target working mode, the target power prediction model is adjusted, and based on the adjusted target power prediction model, the target output power of the target new energy site in the future period is predicted according to the future weather forecast data and the historical output power;
[0324] Send the power prediction model set to the edge server.
[0325] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, database or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processor involved in the various embodiments provided herein may be, but are not limited to, a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic unit, a data processing logic unit based on quantum computing, and the like.
[0326] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0327] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.
Claims
1. A cloud-edge collaborative prediction method for renewable energy power based on automatic machine learning, characterized in that: Executed by an edge server, the method includes: In response to a power forecast demand for a target new energy station, obtaining future numerical weather forecast data of the target new energy station in a future period and historical output power of a historical period corresponding to the future period; According to the missing state of the future numerical weather forecast data and the amount of data of the historical output power, a target power prediction model corresponding to the target new energy site is selected from a power prediction model set; wherein each optional power prediction model in the power prediction model set is trained and issued by a cloud server based on an automatic machine learning algorithm; Determining a target operating mode of the target power prediction model according to the power prediction requirement; wherein the target operating mode includes a prediction service object, a prediction type, and a prediction time scale; According to the target working mode, the target power prediction model is adjusted, and based on the adjusted target power prediction model, the target output power of the target new energy station in the future period is predicted according to the future numerical weather forecast data of the future period and the historical output power of the historical period corresponding to the future period.
2. The method according to claim 1, characterized in that The selecting, according to the missing state of the future numerical weather forecast data and the amount of data of the historical output power, a target power prediction model corresponding to the target new energy station from a power prediction model set comprises: When it is identified that the target new energy station is not a newly built station or a non-capacity expansion station according to the amount of data of the historical output power, the data missing level of the future numerical weather forecast data is determined according to the missing state of the future numerical weather forecast data; According to the data missing level, a target power prediction model is selected from the optional power prediction models belonging to the target new energy site in a power prediction model set.
3. The method according to claim 1, characterized in that According to the missing state of the future numerical weather forecast data and the amount of data of the historical output power, a target power prediction model corresponding to the target new energy site is selected from the power prediction model set, including: When it is identified that the target new energy station is a newly built station or a newly expanded station according to the amount of data of the historical output power, the data missing level of the future numerical weather forecast data is determined according to the missing state of the future numerical weather forecast data; According to the basic information of the target new energy station, a similar station whose similarity with the target new energy station reaches a similarity threshold is selected from other new energy stations; wherein the other new energy stations are stations other than the target new energy station in the station set; According to the data missing level, the optional power of the similar site is selected from the power prediction model set. The target power prediction model corresponding to the target new energy station is selected from the power prediction model.
4. The method according to claim 2 or 3, characterized in that: The target power prediction model after adjustment is used to predict the target output power of the target new energy station in the future period according to the future numerical weather forecast data of the future period and the historical output power of the historical period corresponding to the future period, including: If the data missing level is a scattered missing level, the missing data in the future numerical weather forecast data is supplemented, and based on the adjusted target power prediction model, the target output power of the target new energy station in the future period is predicted according to the historical output power and the supplemented future numerical weather forecast data; If the data missing level is a continuous missing level, the target output power of the target new energy station in the future period is predicted based on the adjusted target power prediction model and according to the historical output power.
5. The method according to claim 3, characterized in that: The selecting, according to the basic information of the target new energy station, a similar station whose similarity with the target new energy station reaches a similarity threshold from other new energy stations, comprises: Obtaining basic information of the target new energy station and basic information of each other new energy station; wherein the basic information includes at least one dimension of latitude and longitude data, capacity data, terrain data, geomorphology data and altitude data; Determine the similarity value between the target new energy station and each other new energy station in each dimension of data according to the basic information of the target new energy station and the basic information of each other new energy station; According to the similarity value between the target new energy station and each other new energy station in each dimension of data, a similar station whose similarity with the target new energy station reaches a similarity threshold is selected from each other new energy station.
6. The method according to claim 1, characterized in that The target power prediction model after adjustment is used to predict the target output power of the target new energy station in the future period according to the future numerical weather forecast data of the future period and the historical output power of the historical period corresponding to the future period, including: If the station type of the target new energy station is distributed new energy, obtaining the grid location of the target new energy station; Based on the adjusted target power prediction model, the target output power of the target new energy station in the future period is predicted according to the historical output power and the future numerical weather forecast data corresponding to the grid location.
7. A cloud-edge collaborative prediction method for new energy power based on automatic machine learning, characterized in that: Executed by a cloud server, the method includes: For each new energy station in the station set, each initial model is trained based on the training sample set corresponding to the new energy station, and an optional power prediction model corresponding to the new energy station is selected from the trained initial models; wherein the training sample set includes sample weather prediction data and sample output power, and different initial models are constructed based on different algorithms; A power prediction model set is constructed according to the optional power prediction models corresponding to each new energy station; wherein the power prediction model set is used for the edge server to select the target power prediction model corresponding to the target new energy station according to the missing situation of future numerical weather forecast data of the target new energy station in the future period and the data amount of the historical output power of the target new energy station in the historical period corresponding to the future period; in the target working mode, the target power prediction model is adjusted, and based on the adjusted target power prediction model, the target output power of the target new energy station in the future period is predicted according to the future numerical weather forecast data of the future period and the historical output power of the historical period corresponding to the future period; The power prediction model set is sent to the edge server.
8. The method according to claim 7, characterized in that The training of each initial model based on the training sample set corresponding to the new energy station includes: Performing data mining on the original weather data in the original data corresponding to the new energy station to obtain sample weather forecast data in the training sample set; Performing multi-stage noise reduction processing on the original output power in the original data to obtain the sample output power in the training sample set; Based on the training sample set, each initial model is trained.
9. The method according to claim 8, characterized in that The data mining process is performed on the original weather data in the original data corresponding to the new energy station to obtain the sample weather forecast data in the training sample set, including: Based on the slicing technology, data cleaning is performed on the original weather data in the original data; Performing correlation analysis on the original output power and the cleaned original weather data to determine a principal component feature set in the original weather data; Verifying the validity of the principal component feature set to obtain a valid feature set; The effective feature set is used as sample weather forecast data in the training sample set.
10. The method according to claim 7, characterized in that The training of each initial model based on the training sample set corresponding to the new energy station includes: For any initial model, obtain the initial hyperparameter space and initial step size corresponding to the initial model; The initial step size is adjusted based on the number of iterations; Based on the adjusted step size, searching for a hyperparameter optimization space in the hyperparameter initial space; In the hyperparameter optimization space, based on the hyperparameter initial value and the Bayesian optimization algorithm, at least one optimized hyperparameter to be verified is obtained; Based on the training sample set corresponding to the new energy station and the super-optimization parameters to be verified, the initial model is subjected to k-fold cross-validation to complete the training of the initial model.
11. The method according to claim 7, characterized in that The step of selecting an optional power prediction model corresponding to the station from each trained initial model includes: Inputting the verification sample set corresponding to the new energy station into each trained initial model to obtain the verification output power output by each trained initial model; Determine the prediction accuracy of each trained initial model according to the actual output power corresponding to the verification sample in the verification sample set and the verification output power output by each trained initial model; According to the prediction accuracy, an optional power prediction model corresponding to the new energy station is selected from each trained initial model.
12. An automatic machine learning new energy power cloud-edge collaborative prediction device, characterized in that: The device comprises: A response module, configured to respond to a power forecast demand for a target new energy station and obtain future numerical weather forecast data of the target new energy station in a future period and a historical output power of a historical period corresponding to the future period; A selection module is used to select a target power prediction model corresponding to the target new energy station from a power prediction model set according to the missing state of the future numerical weather forecast data and the amount of data of the historical output power; wherein each optional power prediction model in the power prediction model set is trained and issued by a cloud server based on an automatic machine learning algorithm; An adjustment module, configured to determine a target operating mode of the target power prediction model according to the power prediction requirement; wherein the target operating mode includes a prediction service object, a prediction type, and a prediction time scale; A prediction module is used to adjust the target power prediction model according to the target working mode, and based on the adjusted target power prediction model, predict the target output power of the target new energy station in the future period according to the future numerical weather forecast data and the historical output power.
13. An automatic machine learning new energy power cloud-edge collaborative prediction device, characterized in that: The device comprises: The training module is used to train each initial model for each new energy station in the station set based on the training sample set corresponding to the new energy station, and select the optional power prediction model corresponding to the new energy station from the trained initial models; wherein the training sample set includes sample weather prediction data and sample output power, different initial models are built based on different algorithms; A building module is used to construct a power prediction model set according to the optional power prediction models corresponding to each new energy site; wherein the power prediction model set is used for the edge server to select a target power prediction model corresponding to the target new energy site according to the missing situation of future numerical weather forecast data of the target new energy site in the future period, and the data amount of the historical output power of the target new energy site in the historical period corresponding to the future period; in the target working mode, the target power prediction model is adjusted, and based on the adjusted target power prediction model, the target output power of the target new energy site in the future period is predicted according to the future numerical weather forecast data of the future period and the historical output power of the historical period corresponding to the future period; The sending module is used to send the power prediction model set to the edge server.
14. An automatic machine learning new energy power cloud-edge collaborative prediction system, characterized in that: include: Cloud servers and edge servers; The cloud server is used to train each initial model for each new energy station in the station set based on the training sample set corresponding to the new energy station, and select an optional power prediction model corresponding to the new energy station from the trained initial models; wherein the training sample set includes sample weather prediction data and sample output power, and different initial models are constructed based on different algorithms; construct a power prediction model set according to the optional power prediction models corresponding to each new energy station; and send the power prediction model set to the edge server; The edge server is used to receive the power prediction model set; in response to the power prediction demand for the target new energy station, obtain the future numerical weather forecast data of the target new energy station in the future period and the historical output power of the historical period corresponding to the future period; according to the missing situation of the future numerical weather forecast data and the data volume of the historical output power, select the target power prediction model corresponding to the target new energy station from the power prediction model set; wherein each optional power prediction model in the power prediction model set is trained and issued by the cloud server based on the automatic machine learning algorithm; according to the power prediction demand, determine the target working mode of the target power prediction model; wherein the target working mode includes the prediction business object, prediction type and prediction time scale; according to the target working mode, adjust the target power prediction model, and based on the adjusted target power prediction model, predict the target output power of the target new energy station in the future period according to the future numerical weather forecast data and the historical output power.
15. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 11 are implemented.
16. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 11 are implemented.
Citation Information
Patent Citations
New energy power station power prediction system based on cloud-side cooperative computing
CN111915053A
Photovoltaic power station power prediction method and system based on recurrent neural network
CN113919232A
New energy power generation power lightweight high-precision cloud prediction system, method and device
CN114462722A
Medium and long term generating capacity prediction method
CN114492922A
Photovoltaic power prediction method and device, electronic equipment and storage medium
CN115660233A
Cited By
Building energy consumption simulation method, system and equipment based on big data and BIM
CN119939745A
Intelligent regulation and control method based on distributed new energy power station
CN120200382A
Power prediction method and device, computer equipment and medium
CN121562450A
Building energy consumption intelligent monitoring and control method
CN121581437A
