A Review of Modern Wind Power Generation Forecasting Technologies

1. Introduction

Of the various sources of renewable energy, wind energy is one of the main types and is growing in use. Worldwide, wind energy reserves are very abundant, and the annual energy that can be developed is about 5.3×10⁷ GWh. The wind power industry is mature, and the methods for renewable energy generation are easy to apply. Wind energy accounted for 6% of global power generation by the end of 2020, with an installed capacity of 743 GW [1]. However, compared with traditional power sources, wind power generation is affected by weather and adjacent terrain environment and is extremely unstable, random, intermittent, and inflexible. Various factors such as wind speed, wind direction, temperature, humidity, atmospheric pressure, and altitude will affect wind power generation. These variables are also interrelated, leading to large fluctuations in wind power, which ultimately makes it difficult to achieve satisfactory results in wind power forecasting. Wind power prediction involves applying state-of-the-art algorithms to the field of wind power generation, such that wind power generation can be better connected to the electricity grid, and key technologies have developed rapidly. In the study of wind power forecasting, wind power has volatility and discontinuity due to the instability of wind itself, which will cause serious difficulties in the scheduling optimization of wind power generation by the electricity grid. Therefore, many efforts and methods have been introduced to solve the wind forecasting problem. Wind power forecasting can be divided into physical methods, statistical methods, artificial intelligence (AI)-based, and deep learning-based methods. Of these methods, the artificial intelligence method can be adaptive and self-learning (e.g., BNN, knowledge graph) in various industries [2,3,4], smart grids [5,6,7], and railway transportation [8], so it is suitable for dealing with the dynamic, nonlinear, and complex characteristics of wind power. Accurate short-term forecasting of wind power is of great significance for alleviating the pressure of power system peak voltage and frequency regulation and wind power connected to the electricity grid.

In order to further improve the accuracy of short-term wind power forecasting, kernel density estimation is used to estimate the probability density function of the random variables required for predictive models to avoid the density leakage problem estimated for probabilistic wind power forecasting (WPF) of a region at both the wind farm and regional levels [9,10,11]. Quantile regression (QR) approximates the conditional probability distribution of a random variable by quantiles. Numerical weather prediction (NWP) data are often used as explanatory variables. Various QR models have been developed for WPPF, such as quantile passive–aggressive regression, regression [12] and curve fitting by weather research and forecasting (WRF), and wind farm parameterization (WFP) as well as quantile regression neural network (QRNN) for regional wind power forecasting (RWPF) [13,14]. In recent years, spatiotemporal forecasting models have been increasingly researched due to their success in improving forecasting accuracy [13]. Given the use of data from different farms and sites to improve the performance of predictive models, spatiotemporal forecasting methods require large amounts of data, which in turn require advanced methods to address the high dimensionality of such situations. A convolution operation to capture the spatial–temporal correlation between neighboring wind farms was based on the novel spatial–temporal wind power predictor (CSTWPP) [15] and a spatiotemporal convolutional network (STCN), each developed separately [16]. New ANN model predictive control-based models [14,17,18,19,20,21,22] have been developed and offered for wind power prediction in microgrid application and use air density and wind speed as input parameters.

The main advantage of ensemble models is their diversity, which allows providing a set of multiple forecasts of the same quantity based on different estimates of initial atmospheric conditions in the WPPF, and ensemble approaches such as the CEEMDAN-IBA-GPR model, multi-feature fusion/self-attention mechanism/graph convolutional network (MFF-SAM-GCN), weighted multivariate time series motifs (WMTSM), and conditional LP (CLP) have been combined with adaptive boundary quantiles (ABQs), wavelet neural network (WNN) trained by the five algorithms [24,25,26,27], data preprocessing (EMD and ICEEMDAN) with parameter optimization [28], and enhanced bee swarm optimization (EBSO) to perform parameter optimization for least squares support vector machine [29] toward probabilistic wind power forecasting, taking full advantage of the most recent information and leveraging the strengths of multiple forecasting models. More recently developed are machine learning methods, which are powerful training algorithms based on artificial intelligence (i.e., neural networks). Due to their high computation intelligence and accuracy, such methods have been widely used in the past few years to improve the accuracy and performance of traditional WPPF models. The machine learning-based wind speed predictions for k-NN and conditional KDE, Adaboost-PSO-ELM, and enhanced bee swarm optimization (EBSO), to perform parameter optimization for least squares support vector machine (LSSVM) [11,26,27,28,30,31] models, were proposed to identify meaningful training data to reduce the volume of modeling data and improve the computing efficiency. They have good generalization ability and robustness and can provide more accurate wind power forecasting. In addition, deep learning is a machine learning concept that provides superior computation performance and flexibility by directly learning the best possible features of raw time series data; for example, the authors of [32,33,34,35,36,37,38,39,40,41,42,43,44] proposed novel data-driven models based on the concepts of deep learning-based convolutional-long short term memory (CLSTM), mutual information, evolutionary algorithm, neural architectural search procedure, and ensemble-based deep reinforcement learning (RL) strategies. The intention of hybrid model forecasting methods [20,24,36,45,46] is to combine different forecasting models to increase the accuracy and precision of forecasts, with their main advantage being that they combine the advantages of each model used to provide the best forecast output. Other statistical analysis methods, such as five minute-ahead wind power forecasts in terms of point forecast skill scores and calibration, 1% point analysis RL-based ESS operation strategy, empirical dynamic modeling (EDM)-based probabilistic forecast, etc. [9,47,48,49,50,51,52,53], were introduced to improve the accuracy of ultra-short-term and short-term wind power forecasts and provide a more reliable basis for wind power grid integration.

To date, several review papers have examined wind power prediction. Wang et al. [54,55] gave an overview of wind power forecasting based on short-term and long-term methods. However, hybrid methods for AI-based wind forecasting have not been studied in detail. A survey by Hanifi et al. [56] was mainly conducted on physical, statistical, and hybrid methods to predict wind power generation. However, the authors explored some AI neural network methods for predicting wind power forecasts, although critical issues and challenges were not explicitly explored. Dhiman and Deb [57] delivered wind speed and wind force forecasting techniques, although deep learning algorithmic methods and their implementation are not covered. Chang [58] proposed a classification of wind power forecasts based on different horizons. Nevertheless, survey of hybrid AI methods and their implementation and limitations were not discussed in detail. Bazionis, I.K. et al. [59] reviewed wind power generation forecasts using various parametric and nonparametric approaches. A classification of wind forecasting methods is given according to timescales, forecasting models, and output data. Hybrid machine learning and deep learning methods have not been fully studied. Furthermore, implementation factors, optimization integration, and hybridization, which are critical issues for hybrid machine learning and deep learning methods, were not outlined. Lipu, M. S. Hossain et al. [60] presented an in-depth investigation of wind power forecasting using artificial intelligence-based hybrid forecasting approaches. Furthermore, various combinations of hybrid AI methods, influencing factors, issues, limitations, and recommendations for achieving wind power forecasting are presented.

In order to meet the needs of large-scale wind power grid integration and further improve the accuracy of short-term wind power forecasting, it is necessary to develop a short-term wind power forecasting model based on advanced hybrid AI algorithms to accomplish accurate, robust, and efficient wind power forecasting. This was achieved here through the following process:

1. This paper begins by summarizing the time resolution, model type, accuracy, and parameters of current advanced wind power forecasting technologies and determines the classifications, advantages and disadvantages, and contributions of the various wind power forecasting models.

2. These models have different predictive capabilities, and the weights of each model are updated in real time to improve the comprehensive predictive capabilities of the models, which have good application prospects in wind power forecasting.

3. Case studies and examples in the literature of accurate ultra-short-term and short-term wind power forecasting predictions with uncertainty and stochasticity are reviewed and analyzed.

Finally, the conclusion is drawn, and existing issues in the methodologies outlined. Future prospective research directions are presented.

2. Review of research status

European and American countries, such as Denmark, the United States, and Spain, have developed relatively early in terms of studying wind power generation, and many advanced results have been obtained as the basis for the maturation of research on wind power forecasting systems [61,62]. Based on meteorological information, they have built a relatively complete wind power forecasting system with the NWP system as the core. Prediktor is a prediction system developed by Denmark's Risø DTU National Laboratory for Sustainable Energy and put into use in 1994 [63]. Denmark is in a leading position in the development of wind power forecasting. For example, the WPPT forecasting system uses a combination of adaptive least squares and exponential forgetting algorithms (least squares and exponential forgetting algorithms), which can provide forecasts ranging from 0.5 to 36 hours [64]. The Zephyr forecasting software developed by the Risø DTU National Laboratory is very popular in Denmark. It combines physical models with adaptive least squares and exponential forgetting algorithms to provide forecasts from 0 to 9 hours and 36 to 48 hours [65]. It uses a physical model and considers the impact of wind turbine wakes. By combining statistical methods with physical methods, the eWind system developed by American Truewind Company adopts a combination of physical and statistical forecasting methods, which can upload real-time information and online in-depth analysis and has the ability to accurately predict the next 48 hours [66]. The WPMS forecasting system developed in Germany is the most widely used forecasting software at present [67]. The system is combined with neural network on the basis of NWP forecasting, and further improves the forecasting accuracy of wind power. Also in Germany, ISET has developed the forecasting system AWPT [68], which was put into operation in 2001 and uses the method of combining NWP and neural networks mainly in 1- to 8-hour forecasting. The following year, the University of Oldenburg in Germany developed the Previento system, which added typical physical models to the prediction system and could accurately obtain 2-day prediction results [69]. The Siperolico forecasting software developed by Carlos III University in Spain, HIRPOM [70] in Ireland, and the LocalPred model of the Renewable Energy Operation Centre (CORE) in Spain use both statistical and physical models [71]. The ANEMOS project has a total of 23 institutions from 7 countries, including Ireland, France, and Spain, participating in the research and development, which can predict the wind power of large-scale offshore and land wind farms. Multiple NWP models are used in the ANEMOS project, so the local meteorological department is required to provide numerical weather prediction data [72]. After processing, physical methods and statistical methods are used to make predictions, whose accuracy can reach about 10%.

Different prediction algorithms are selected for prediction according to differences in regions, weather, and climate types. In addition to using physical methods for prediction, the system can perform statistical analysis of historical wind power data, further improving the prediction accuracy. In November 2008, the WPFS system was developed by China [73]. After a series of successful test experiments, the system became the first mature wind power forecasting system in China. The system can effectively predict the short-term wind power within three days and the ultra-short-term wind power within four hours. The wind power generation forecasting systems currently used around the world are summarized as shown in Table 1.

2.1. Reviews for technologies and applications

In recent years, relevant scholars have conducted theoretical research and practical simulation. The prediction type of wind speed has different definitions according to the length of the cycle, and different researchers have different classifications as shown in Table 2, mainly including ultra-short-term, short-term, medium-term, and long-term prediction.

Long-term forecasting based on "month" and "year" is mainly used in the design of wind farm operation plans and the evaluation of wind power resources for the planning of wind farms. The medium-term forecast is usually used to predict the sampling points in the next few days or "week" and is mainly used for troubleshooting and maintenance of wind power equipment in the power grid. Short-term and ultra-short-term forecasting is based on "hour" or "minute", which is mainly used to adjust the reserve capacity of the power system and economic dispatching to reduce the instability of the system caused by wind power connected to the electricity grid, so as to conduct effective grid dispatching. According to different modeling methods, wind power generation forecasting can be divided into physical methods, statistical methods, artificial intelligence methods, and deep learning methods. Depending on the different prediction objects, it can be divided into indirect corresponding wind speed prediction and direct corresponding power prediction, as shown in Figure 1. According to the forecasting model, wind power forecasting can be divided into direct forecasting and indirect forecasting. Direct prediction refers to the establishment of a corresponding mathematical model based on the historical wind power time series of the wind farm to predict the future wind power [74].

According to the prediction principles, wind power prediction can be divided into physical methods, statistical analysis methods, artificial intelligence methods, methods based on deep learning, and combined prediction models. In physical methods, the relatively rough forecast value output by the numerical weather forecast system is analyzed for making predictions based on the physical information around the wind farm and meteorological information such as weather and temperature. The advantage of physical methods is the lack of requirement for supporting historical wind farm power data. The disadvantage is that they are very sensitive to initial parameters, such as terrain description information. Inaccuracy in the initial parameters are inaccurate will cause large prediction errors. Statistical analysis requires a large amount of history of wind power or wind speed for statistical analysis, such as using Markov chain [12,15,75], regression analysis [29,30,35,41,47,76], Kalman filtering [77], and ARMA [29,78] models to find the laws contained in historical data for prediction. The advantage of the statistical method is that under the premise of sufficient historical data, the forecast error can be minimized in theory, and the forecast accuracy is high, but the disadvantage is that a large amount of historical data is required for support. The deep learning method is an emerging prediction method that can use artificial intelligence to establish an accurate model describing the nonlinear relationship between input and output. It can predict from the essence of wind energy, thereby improving the prediction accuracy. Common methods include neural network [79], wavelet analysis [44,80], and support vector machine [73,78,79,81]. Hybrid predictive models of artificial intelligence methods are becoming increasingly popular, not only increasing the complexity of algorithms but also enhancing the forecasting of wind power generation. Typically, hybrid predictive models are designed by combining two or three deep learning techniques or combining optimization algorithms with AI methods. This addresses the aforementioned shortcomings of a single predictive model by finding optimal features, hyperparameters, and training algorithms. The review focuses on wind power generation forecasting for time resolution, and the model type, accuracy, and parameters used are presented in Table 3, which shows that over 70% of wind power prediction literature uses RMSE and MAE methods to evaluate errors. These results indicate that when wind speed changes rapidly and there is a slight error, even if the wind speed value is small, the error value will be amplified and thus, there are many methods for error determination in wind power prediction. In accuracy of Table 3, we have compiled and proposed many methods for computing errors, such as RMSE, MAE, MASE, NRMSE, MAD, MRE etc. and try to present the error values as fully as possible in this review paper, where future wind power prediction research needs to improve and evaluate.

2.2. Problem

In order to using physical models for calculation or statistical methods for simulation, the wind speed and wind direction of the wind farm in a short period of time in the future are predicted. These are the highly relevant meteorological parameters with wind power to predict the ultra-short-term or short-term performance of the wind farm in the future. The output power provides a basis for the power sector to execute power generation scheduling, modeling, and expansion planning. The fluctuation of the wind is relatively large, and there are often jumps in time to cause high probability of random uncertainty. As the capacity of newly installed and operated wind turbines continues to grow, they will occupy an increasing proportion in the electrical grid. However, the penetration power of wind power generation cannot exceed its limit value. If exceeded, the wind turbine units connected to the power grid will pose a huge threat to the power grid, making the power system unable to perform normal, stable, and safe operation. In order to solve the above problems, it is necessary to scientifically predict the power output of wind farms according to the changing trend of wind, wind speed, and wind direction, so as to improve the controllability of wind power generation. If the output power of wind power generation can be more accurately predicted through parameters such as wind change trends, wind speed, and wind direction, then the predicted data will be uploaded to the power dispatching center. The power dispatch center can scientifically and efficiently control power generation and distribution based on these data. Reasonable arrangements can fundamentally reduce the impact of wind power generation on the electrical grid and greatly increase the grid connection rate of wind power generation. Accurate wind power prediction solves the problem of grid connection to reduce the operating cost of wind farms. Therefore, wind power prediction technology has attracted the global attention of wind power fields, scholars, enterprises, and departments and is of great significance to the development of wind power.

Most prediction systems combine physical and statistical concepts, and their accuracy is limited by the numerical weather prediction model (NWP). When the forecast time exceeds 6 hours, the numerical weather prediction (NWP) should reduce the temporal and spatial scale of the wind field to convert the local wind speed into electrical energy and then estimate the power of the entire region for a single wind field. The prediction error is about 10-15% of the root mean square error (RMSE) for the capacity of the whole wind farm. However, with the increasing capacity of wind turbines, the requirements for the accuracy of wind power generation prediction will be stricter. The traditional math equation for calculating power generation cannot directly reflect the rapid change of wind speed, and there is always a great error between the calculated value and the actual value.

2.3. Comparative study of the reviewed WPPF models and methodologies

In the past decade, the research on wind power generation prediction has become increasingly popular. Most models use numerical weather prediction (NWP) and on-site measurement data (SCADA) as the basis, read the data of monitoring points, and then use the obtained wind speed and output data to predict wind power. However, due to the confidentiality of data sources, most prediction models rarely have complete theoretical basis and historical data; therefore, the research is limited to a single site and region. A single prediction may be reflected by higher forecast accuracy. Once considering multiple regions or large-scale spatial prediction in scattered meteorological stations, the results must be questionable. In addition to the increase in offshore wind turbines and the large amount of investment in private wind farms, persuading private wind farms to provide substantial wind turbine information (wind power generation, operating conditions, etc.) is a major challenge. Therefore, it is necessary to devote efforts to the prediction of large spatial scales. In this paper, the contribution of the current advanced wind power forecasting technology is summarized, outlining the distinct advantages and disadvantages of the various wind power forecasting models. These forecasting models have different forecasting capabilities, update the weights of each model in real time, improve the comprehensive forecasting capability of the model, and have good application prospects in wind power load forecasting. Finally, this paper remarks on the contributions, advantages, disadvantages and approaches of reviewed works in terms of wind power forecasting in Table 4. In previous wind power prediction studies, most researchers used past meteorological data for evaluation. However, we were able to obtain more data such as satellite data, future meteorological data, etc. due to the advanced information techniques. In Table 4, more than 50% of the literature on wind power prediction used more input data than previous studies, without optimal feature based data preprocessing; Meanwhile, the prediction structure methods used are becoming increasingly complex, resulting in longer computation times. Some papers in the review literature also propose to preprocess historical wind data to reduce training times, thereby achieving effective data screening and improving the accuracy of wind power prediction.

3. Opportunities and Challenges

Our research team used statistical methods to predict wind power generation in the past and achieved the purpose of predicting wind power generation through wind speed prediction. Due to the large amount of historical data, we currently use the analysis and selection of data features to reduce the training sample space and improve the extraction of more useful data from the database, thereby reducing the amount of data, shortening the running time, and obtaining good prediction results; then, a wind power generation model is established based on support vector machines as a predictive method suitable for nonlinear regression analysis. Its internal parameters will affect the accuracy of regression analysis. Therefore, the bee colony algorithm is used to better solve the parameter values. In order to reduce the prediction error, we aim to minimize the error and use the bee colony optimization algorithm to solve the parameter setting problem of the support vector machine, which not only increases the integrity of the prediction model but also improves the prediction accuracy. We use meteorological observation stations for long-term recording and monitoring to obtain relevant wind speed and power generation data, and data regression technology is used to preprocess the input data to remove parameters that reduce the predicted value, thereby reducing network input parameters. The test results validate the effectiveness and accuracy of the proposed prediction method.

3.1. Bee Colony Algorithm Combined with Data Regression Support Vector Machine

Support vector machine is a pattern identification algorithm of artificial intelligence. It is a new machine learning method based on statistical learning theory and shows many unique advantages in solving small sample, nonlinear, and high-dimensional pattern recognition problems. It has been applied to practical problems such as wind power prediction, handwriting recognition software, face recognition, and image classification. In addition, the bee colony algorithm is a new optimization algorithm that can accurately and quickly solve numerical optimization problems. It has the advantages of simple concept, easy implementation, and fewer control parameters.

The wind power forecasting process is mainly divided into three parts, and the flow chart is shown in Figure 2(a).

Part I: Apply data regression analysis to select modeling data.

Part II: Applying the bee colony algorithm to solve the optimal setting parameters of the support vector machine, the EBSO_LSSVM system architecture is shown in Figure 2(a).

Part III: Performing wind speed prediction using EBSO_LSSVM.

As support vector machine parameter selection has a great influence on the regression analysis, the parameter selection will often vary in different cases; therefore, the proposed method uses data regression support vector machine as the main body combined with the enhanced bee swarm algorithm, which is referred to as EBSO_LSSVM. The enhanced bee swarm algorithm was used to find better setting parameters for γ and σ in the support vector machine resulting in better training parameters, thus improving on the traditional shortcomings of unclear support vector machine parameter selection. EBSO_LSSVM optimization algorithm has unique advantages in dealing with small sample problems as follows:

(a) Use the algorithm to find the best modeling parameters: EBSO_ LSSVM uses EBSO to adjust parameters to achieve the best modeling of LSSVM. This method effectively improves the accuracy of LSSVM prediction. However, EBSO uses to select LSSVM parameters for each prediction. Each prediction of a time interval requires the construction of an LSSVM, which will lead to excessively long modeling operation time, so DR-SVM is proposed.

(b) The method of characterizing the training data is modeling: DR-SVM is a self-organizing map (SOM), in which a database is first used to cluster and quantify the relevant data (feature clustering), and the whole clustered data are drawn in a normal way to avoid influences from differences in too large or small data values. Because modeling after clustering and quantification will have strong resilience to prediction, the performance will be better than that of the untreated [82].

(c) Modeling method combining algorithm and data characteristics: The research team recommends DR_EBSO_Combining LSSVM, where it is not necessary to build a prediction model every time resolution, because DR has greater flexibility, which means DR_EBSO_LSSVM can continuously predict for long prediction times.

Figure 2(a) presents the SVM by using the radial basis function (RBF) kernel. DR is a self-organizing map (SOM), as shown in Figure 2(b), and the algorithm of SOM can be self-defined according to the use environment and meet the purpose of data visualization. The sample dataset containing n-dimensional vector is taken as the matrix structure in two-dimensional space. Under the condition that the topological structure between the sample points of the original dataset remains unchanged, a spatial structure with a dimension of no more than n by correspondence is obtained. Each point in the spatial structure contains a weight vector, and SOM can be regarded as a sample point in the dataset. The spatial structure obtained by nonlinear regression. In this paper, the layer Z1 of the four main datasets, wind, temperature, humidity, and rainfall, is constructed in the first layer (constructed in the two-dimensional topological structure space 2x3), and the secondary topological structure of the four main datasets (layer Z21~layer Z24, constructed in the two-dimensional topological structure space 2x3) is constructed in the second layer, as shown in the following figure. For the Z21-Z24 topology of the second layer, T and N3, are the prediction time and time interval, respectively. T-N3 represents the historical meteorological data value in the past T-N3, and T+N3 represents the weather forecast data value in the future T+N3.

3.2. Wind Data Diversity

The challenge of wind power forecasting models is the variability in wind data and the datasets from different historical or meteorological data. The fluctuation and intermittency of wind energy lead to uncertainty in the output power of wind farms, which will have an adverse impact on the power system. In particular, the large-scale grid connection of wind power will inevitably bring difficulties to grid dispatching, thereby reducing the reliability of the grid. Accurate wind power forecasting is crucial for wind power grid integration and the stable operation of power systems. Therefore, the wind power forecasting models should be verified under various data preprocessing techniques in further research.

(1)

Wind speed prediction process

The wind is highly intermittent, so forecasting is an extremely difficult issue, but this is inseparable from the preliminary work of wind power forecasting modeling. Its input variables are temperature, humidity, and wind speed, the output variable is the wind speed value in the next period, and support vector machine is used to make one-hour wind speed prediction and one-day wind speed prediction. Figure 3 shows the establishment of the wind speed prediction process in this paper.

(2)

Establishment of power generation model

In the value of actual power generation, there are often points with large fluctuations in power generation, or points with abnormal values caused by improper instrument measurement. This "noise" power generation will be subjected to two rounds of regression data analysis in this study. In exploration, the first step is to establish cluster points for all historical data, and the second is to filter data based on the first cluster point, deleting the 20% data that deviates from cluster power generation in each interval. They will be excluded in the reference data for data regression analysis, and a more complete cluster point is appropriately modeled using 80% of the data. The process is shown in Figure 4.

3.3. The state-of-the-art approaches for short-term wind power forecasting

The prediction effects of different predictive models have their own advantages and disadvantages. The hybrid prediction method involves optimizing and combining the results of data processing of different models according to a specific strategy, so as to obtain the better wind power prediction results and ultimately achieve the purpose of improving the accuracy of wind power prediction. In the final analysis, the combined forecasting method is used to optimize the forecasting results, and as long as the time series scale of the power forecasting output corresponds to the combined forecasting of the model, there is no limit to the relevant algorithms used for each forecast. At present, due to the large amount of data that can provided by wind farm numerical weather prediction (NWP), combining NWP information to improve the accuracy of wind power forecasting has become the main research direction in hybrid wind power forecasting.

The work described in this paper is shown in Figure 5, and the short-term wind power forecasting model is discussed in depth. The latest approaches of short-term wind power forecasting in the past three years are reviewed to provide an important reference in wind power grid integration. In order to improve the accuracy of wind power forecasting, this paper gives a detailed overview of the contributions, advantages, and disadvantages of various delivered wind power forecasting models and future research works. These advanced forecasting methods can be approximately classified into kernel density estimation, quantile regression methods (QR), artificial intelligence/neural networks (NN), ensemble methods, spatiotemporal forecasting, machine learning, deep learning, hybrid model forecasting, and other statistical analysis methods. These proposed novel short-term wind power forecasting models provide very useful information for power system operation and control with high renewable energy penetration.

In recent years, various research institutions and scholars have adopted different state-of-the-art approaches to improve the problems of power fluctuation and randomness in wind power forecasting as well as possible errors and omissions in original data, and certain advancements have been made, but there are still some issues that need to be urgently solved. First of all, in the future, the sample space can be further expanded, and the dimension of data samples can be increased to predict wind data diversity. According to the wind power data with different characteristics, the prediction model is further optimized to increase the applicability of the model. Secondly, according to the characteristics of the existing hybrid model, the parameter optimization method is further improved to ensure that the prediction model has high prediction accuracy at different time sampling rates, making it suitable for different prediction occasions.