New energy storage consumption capability index prediction method

Abstract

The invention provides a new energy storage consumption capability index prediction method, and relates to the field of power system operation, and the method comprises the steps: building a nonlinear mapping relation between solar radiation intensity, photovoltaic module conversion efficiency, a temperature coefficient, an aging attenuation rate factor and power generation efficiency for photovoltaic power generation, so as to dynamically correct a photovoltaic power generation potential output prediction value; based on a potential output prediction result of photovoltaic and wind power generation, solving an energy storage capacity configuration and operation strategy through an intelligent optimization algorithm to maximize the benefit of an energy storage system, and obtaining a dynamic prediction value of an energy storage consumption capability index; and combining prediction results of the statistical model based on the historical data, the dynamic model based on the physical mechanism and the data-driven model based on machine learning by adopting an adaptive weighted fusion strategy. According to the method, accurate prediction of potential output of new energy power generation and dynamic evaluation of the energy storage absorption capability are realized, and effective technical support is provided for new energy grid connection and optimal configuration of an energy storage system.

Description

Technical Field

[0001] The present invention relates to the field of power system operation, and in particular to a method for predicting a new energy storage and absorption capacity index. Background Art

[0002] In the prediction of the new energy storage and absorption capacity index, a key technical issue is how to accurately predict the potential output of new energy power generation. Taking photovoltaic power generation as an example, the factors that affect its potential power generation include solar radiation intensity, photovoltaic module conversion efficiency, module temperature coefficient, module aging attenuation rate, etc. These factors themselves are also affected by many uncertain factors, such as weather conditions, atmospheric transparency, cleanliness of photovoltaic modules, and module degradation rate. Similarly, the potential wind energy resources of wind power generation are also affected by factors such as wind speed, wind direction, air density, and aerodynamic characteristics of wind turbine blades. The uncertainty and volatility of these factors have brought great challenges to the prediction of the potential output of new energy power generation.

[0003] How to establish an accurate and robust prediction model for the potential output of renewable energy power generation under various uncertain factors, and organically combine it with the energy storage system's absorption capacity assessment model to form a complete energy storage absorption capacity index prediction framework is a key technical problem that needs to be solved urgently. This requires in-depth analysis of the interaction and transmission mechanism between various influencing factors, the development of advanced data fusion and mining algorithms, the construction of multi-time scale and multi-spatial resolution prediction models, and the verification and optimization of the models through a large amount of measured data, and finally the formation of a practical and reliable energy storage absorption capacity index prediction technology system, which provides important decision-making support for the efficient use of renewable energy and the safe and stable operation of the power system. Summary of the invention

[0004] The present invention provides a prediction of the new energy storage and consumption capacity index, which mainly includes:

[0005] Obtain historical meteorological data and geographic information data, and based on the historical meteorological data and geographic information data, establish a multivariate regression model of key factors and weight coefficients affecting the potential output of renewable energy power generation, use the Monte Carlo random sampling method to generate a large number of input samples, substitute the input samples into the renewable energy power generation potential output calculation model to obtain the probability distribution of potential output, perform wavelet analysis on the probability distribution to obtain the output fluctuation characteristics under different time scales, and perform abnormal fluctuation warning based on the output fluctuation characteristics; for photovoltaic power generation, establish a nonlinear mapping relationship between solar radiation intensity, photovoltaic module conversion efficiency, temperature coefficient, aging attenuation rate factors and power generation efficiency to dynamically correct The predicted value of potential output of photovoltaic power generation; for wind power generation, a computational fluid dynamics model is used to simulate the wind field distribution, obtain the wind speed and wind direction parameters at the location of the wind turbine, and determine the potential output of wind power in combination with the wind turbine power curve, and make corrections based on the abnormal fluctuation warning results; based on the potential output prediction results of photovoltaic and wind power generation, the energy storage capacity configuration and operation strategy are solved through an intelligent optimization algorithm to maximize the benefits of the energy storage system, and the dynamic prediction value of the energy storage absorption capacity index is obtained; and an adaptive weighted fusion strategy is used to combine the prediction results of the statistical model based on historical data, the dynamic model based on physical mechanisms, and the data-driven model based on machine learning to predict the energy storage absorption capacity index.

[0006] Furthermore, the Monte Carlo random sampling method is used to generate a large number of input samples, including: based on the uncertainty distribution function of the key factor, using the Monte Carlo random sampling algorithm to generate a large number of random input sample data that conform to the distribution characteristics, and preprocessing the random input sample data.

[0007] Furthermore, for the preprocessed data, the chi-square test and Kolmogorov-Smirnov test are used to determine whether each factor obeys the normal distribution or Weibull distribution. If the factor obeys a known distribution type, the maximum likelihood estimation method is used to estimate the corresponding distribution parameters. According to the probability density function obtained by the estimation, a large number of random samples that obey the distribution are generated by Monte Carlo simulation for subsequent risk assessment.

[0008] Furthermore, the wavelet analysis of the probability distribution is performed to obtain the output fluctuation characteristics at different time scales, including: discretizing the probability density function, performing wavelet packet decomposition as the input signal of the wavelet transform, obtaining wavelet coefficients at multiple time scales, determining the abnormal fluctuation threshold of each scale based on historical data, judging whether the wavelet coefficient exceeds the threshold to obtain abnormal fluctuation, and summarizing the abnormal fluctuation to construct a feature vector.

[0009] Furthermore, the random input sample data is preprocessed, including: removing outliers and missing values, and normalizing the data using the maximum and minimum normalization method to map the data to [0, 1].

[0010] Within the interval, the preprocessed input sample data is obtained, and according to the pre-established physical mechanism model of the potential output of new energy power generation, the preprocessed input sample data is used as training data, and the mean square error loss function is minimized by the gradient descent algorithm and the back propagation algorithm to optimize the parameters of the physical mechanism model to obtain the trained physical mechanism model, and the preprocessed input sample data is substituted into the trained physical mechanism model and the data-driven model respectively to calculate the potential output value corresponding to each input sample, and the weighted average of the prediction results of the two models is performed to obtain the final potential output prediction value.

[0011] Furthermore, for photovoltaic power generation, a nonlinear mapping relationship is established between solar radiation intensity, photovoltaic module conversion efficiency, temperature coefficient, aging attenuation rate factors and power generation efficiency to dynamically correct the predicted value of photovoltaic power generation potential output, including: using a deep learning algorithm to establish a nonlinear mapping relationship between the factors and power generation efficiency, and correcting the predicted value according to the abnormal fluctuation warning results.

[0012] Furthermore, a probability density function of the data is obtained, and the probability density function is discretized by a histogram method to obtain a discretized probability density function, including: using the discretized probability density function as an input signal of a wavelet transform, performing multi-scale decomposition by a Daubechies wavelet to obtain wavelet coefficients at multiple time scales, determining an abnormal fluctuation threshold of the wavelet coefficients at each time scale according to preset historical data, if the wavelet coefficient at a current moment exceeds the abnormal fluctuation threshold, then it is determined that an abnormal fluctuation exists at that moment, the statistical features of the abnormal fluctuation are extracted, an abnormal fluctuation feature vector is constructed, a random forest algorithm is used to classify the abnormal fluctuation feature vector, and an anomaly detection model is obtained, and if the abnormal fluctuation feature vector of the new input data is determined to be abnormal by the anomaly detection model, then an early warning is triggered.

[0013] Furthermore, for wind power generation, a computational fluid dynamics model is used to simulate the wind field distribution, the wind speed and wind direction parameters at the location of the wind turbine are obtained, the potential wind power output is determined in combination with the wind turbine power curve, and corrections are made according to the abnormal fluctuation warning results, including: based on the three-dimensional model of the wind field, a computational fluid dynamics method is used to simulate the speed and direction distribution of the wind field, the wind speed and wind direction data at the location of the wind turbine are obtained, the wind speed data is input into the power curve of the wind turbine to calculate the theoretical power generation of the wind turbine at the wind speed, and corrections are made according to the abnormal fluctuation warning results.

[0014] Furthermore, the potential output prediction results based on photovoltaic and wind power generation are used to solve the energy storage capacity configuration and operation strategy through an intelligent optimization algorithm to maximize the benefits of the energy storage system, and obtain a dynamic prediction value of the energy storage absorption capacity index, including: according to the potential output prediction results of photovoltaic and wind power generation, an intelligent optimization algorithm is used to maximize the benefits of the energy storage system while satisfying the constraints of new energy absorption and system stability, and obtain a dynamic prediction value of the energy storage absorption capacity index.

[0015] Furthermore, the adaptive weighted fusion strategy is used to combine the prediction results of the statistical model based on historical data, the kinetic model based on physical mechanism and the data-driven model based on machine learning to predict the energy storage capacity index, including: dynamically adjusting the weight coefficient of each sub-model according to the prediction error to improve the overall accuracy and robustness of the energy storage capacity index prediction.

[0016] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0017] The present invention discloses a method for predicting the energy storage and absorption capacity index of new energy. The method first establishes a multivariate regression model to determine the key factors affecting new energy power generation and their uncertainty distribution, and generates a potential output probability distribution through Monte Carlo sampling. Wavelet analysis is performed on the distribution to achieve multi-scale decomposition and set a threshold to warn of abnormal fluctuations. For photovoltaic power generation, a deep learning algorithm is used to establish a nonlinear mapping relationship to dynamically correct the prediction value. For wind power generation, a computational fluid dynamics model is used to simulate the wind field distribution and the potential output is determined in combination with the wind turbine power curve. Based on the above prediction results, the energy storage capacity configuration and operation strategy are solved by an intelligent optimization algorithm to obtain a dynamic prediction value of the energy storage absorption capacity index. Finally, an adaptive weighted fusion strategy is used to combine the prediction results of multiple models to improve the overall prediction accuracy. The present invention realizes the accurate prediction of the potential output of new energy power generation and the dynamic evaluation of energy storage absorption capacity, and provides effective technical support for the grid connection of new energy and the optimal configuration of energy storage systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 The present invention is a flowchart of a method for predicting a new energy storage and absorption capacity index. DETAILED DESCRIPTION

[0019] The following will describe the technical solutions in the embodiments of the present invention in detail in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention.

[0020] like Figure 1 In this embodiment, a new energy storage capacity index prediction method may specifically include:

[0021] Step S101, based on historical meteorological data and geographic information data, a multivariate regression model is established using a machine learning algorithm to obtain key factors and weight coefficients that affect the potential output of renewable energy power generation, and to determine the uncertainty distribution function of each factor. The uncertainty distribution function is used to describe the random variation characteristics of each influencing factor and provide a probability distribution basis for subsequent Monte Carlo sampling.

[0022] According to historical meteorological data and geographic information data, machine learning algorithms such as random forests are used to evaluate the importance of features and screen out the key factors that affect the potential output of renewable energy power generation. For the selected key factors, their historical data are obtained, and the probability distribution type of each factor is determined by statistical test methods. If it obeys the normal distribution, the mean and variance parameters are estimated by the maximum likelihood estimation method; if it obeys the Weibull distribution, the scale parameter and shape parameter are estimated to obtain the probability density function of each key factor. According to the estimated probability density function, the Latin hypercube sampling method is used to generate the Monte Carlo sample set, and the sample size can be set to 1000. The generated sample set is input into the subsequent multivariate linear regression model to obtain the predicted value of renewable energy power generation under each sample. Based on the key factors screened in the first step and the sample set obtained by Monte Carlo simulation, a multivariate linear regression model is established to fit the quantitative relationship between renewable energy power generation and each key factor, and the regression coefficient reflects the influence weight of each factor. Substitute the Monte Carlo sample set into the regression model to obtain the predicted value of renewable energy power generation under each sample, and draw the probability distribution histogram of the predicted power generation. The power generation at different quantiles (such as P50, P90, etc.) is calculated as an evaluation indicator of the new energy power generation potential and provides a reference for decision-making on new energy projects.

[0023] Exemplarily, there are many factors that affect the potential of renewable energy generation, such as meteorological factors such as light intensity, wind speed, temperature, humidity, air pressure, and geographic information factors such as terrain, altitude, slope, and aspect. In order to accurately evaluate the potential of renewable energy generation, it is necessary to screen out key influencing factors. The random forest algorithm can evaluate the importance of features. For example, by collecting 10 years of historical meteorological data and geographic information data in a certain area, as well as renewable energy power generation data in the corresponding time period, a random forest model is constructed. After the model training is completed, the importance ranking of each feature can be obtained. Assume that the ranking results show that light intensity, wind speed, and temperature are the three most critical factors affecting the potential of renewable energy generation in the region. After screening out the key factors, it is necessary to determine the probability distribution type of these factors. Obtain historical data of these key factors. For example, collect daily light intensity, wind speed, and temperature data for the past 10 years in the region. Use statistical test methods, such as the Kolmogorov-Smirnov test or the Anderson-Darling test, to determine the probability distribution type of each factor. The hypothesis test results show that light intensity follows a normal distribution, wind speed follows a Weibull distribution, and temperature follows a normal distribution. For light intensity and temperature that follow a normal distribution, the maximum likelihood estimation method can be used to estimate their mean and variance. For example, the sample mean of light intensity calculated based on historical data is 800W / m 2 , the sample variance is 10000(W / m 2)2; the sample mean of temperature is 20℃, and the sample variance is 9℃2. For the wind speed that obeys the Weibull distribution, its scale parameter and shape parameter can be estimated. For example, according to the historical data, the scale parameter of the wind speed is 6m / s, and the shape parameter is 2. So far, the probability density functions of the three key factors have been obtained. After obtaining the probability density function, the Latin hypercube sampling method can be used to generate the Monte Carlo sample set. Set the sample size to 1000. For example, according to the probability density functions of light intensity, wind speed and temperature, 1000 random samples that obey their respective distributions are generated to form a sample set of 1000 rows and 3 columns. Each row represents a possible scenario, containing the values ​​of the three factors of light intensity, wind speed and temperature. Based on the selected key factors (light intensity, wind speed, temperature) and the 1000 samples generated by Monte Carlo simulation, a multivariate linear regression model can be established. This model is used to fit the quantitative relationship between renewable energy power generation and light intensity, wind speed and temperature. For example, suppose the established regression model is: power generation = a*light intensity + b*wind speed + c*temperature + d, where a, b, c, and d are regression coefficients, reflecting the weight of each factor on power generation. By fitting historical data, the specific values ​​of these coefficients can be obtained. Substituting each sample in the Monte Carlo sample set into the established multivariate linear regression model, the predicted value of new energy power generation under 1000 samples can be obtained. Plotting these 1000 predicted values ​​into a probability distribution histogram can intuitively show the distribution of new energy power generation potential. According to the probability distribution histogram, the power generation at different quantiles (such as P50, P90, etc.) can be calculated. P50 means that there is a 50% probability that the power generation will reach or exceed this value, and P90 means that there is a 90% probability that the power generation will reach or exceed this value. These quantile values ​​can be used as evaluation indicators for new energy power generation potential and provide reference for new energy project decision-making. For example, if the power generation corresponding to P90 is lower than the expected value, the feasibility of the project may need to be re-evaluated. Doing so can quantify risks and guide investment decisions for new energy projects more scientifically.

[0024] Step S102, based on the historical meteorological data and geographic information data, a statistical test method is used to determine the probability distribution type of each factor. If it obeys a normal distribution, the mean and variance parameters are estimated using the maximum likelihood estimation method. If it obeys a Weibull distribution, the scale parameter and shape parameter are estimated to obtain the probability density function of each key factor.

[0025] Historical meteorological data and geographic information data were obtained, and the data were preprocessed, including using interpolation to handle missing values, using box plots and other methods to identify and remove outliers, and using minimum-maximum standardization and other methods to standardize the data. For the preprocessed data, statistical test methods such as chi-square test and Kolmogorov-Smirnov test were used to determine whether each factor obeyed common probability distribution types such as normal distribution and Weibull distribution. For factors that obey known distribution types, the maximum likelihood estimation method was used to estimate the corresponding distribution parameters. Taking the normal distribution as an example, the log-likelihood function was constructed based on the sample data, and the maximum likelihood estimates of the mean and variance were obtained by solving the log-likelihood equation group. Similarly, for factors that obey the Weibull distribution, the scale parameter and shape parameter were estimated by constructing and solving the log-likelihood equation group. The estimated parameters were substituted into the corresponding probability density function expression to obtain the probability density function of each key factor. On this basis, the characteristic quantities such as the mean and variance of the probability density function are calculated by numerical integration and other methods, and the commonly used quantiles are obtained by solving the inverse function of the probability distribution function, which provides a basis for subsequent analysis. Using the estimated probability density function, a large number of random samples that obey the distribution are generated by Monte Carlo simulation. Specifically, the probability distribution function is first obtained based on the probability density function, and then a random sample that obeys the target distribution is obtained by generating uniform random numbers on the interval [0, 1] and substituting them into the inverse function of the probability distribution function. Repeat the above process to obtain a sufficient number of random samples. Based on the generated random samples, the risk level and uncertainty of the system in different scenarios are evaluated by statistical analysis and hypothesis testing. For example, the expected value and volatility of the system risk are estimated by the mean and variance of the sample; the risk threshold at a given confidence level is estimated by the sample quantile; and the difference in risk levels under different scenarios is judged by hypothesis testing. According to the results of risk assessment, the system design and operation strategy are optimized and effective risk management measures are formulated. For factors that do not obey known distribution types, non-parametric estimation and other methods can be used to directly estimate their probability density functions based on sample data. For example, using methods such as kernel density estimation, by selecting appropriate kernel functions and bandwidth parameters, fitting the empirical distribution of the sample, a non-parametric estimate of the probability density function can be obtained. On this basis, various characteristic quantities of the probability density function of the non-parametric estimate can be further calculated and used for subsequent risk assessment and decision analysis.

[0026] For example, meteorological data and geographic information data are important foundations for assessing the potential of renewable energy generation. After obtaining these data, they need to be preprocessed first to ensure the quality and reliability of the data. For example, collect 10 years of historical meteorological data in a certain area, including daily light intensity, wind speed, temperature, humidity, etc., as well as geographic information data such as altitude, slope, aspect, etc. In the process of data collection, data may be missing, such as the meteorological data of some days is not recorded due to equipment failure. At this time, interpolation can be used to fill the missing values. For example, linear interpolation, spline interpolation and other methods can be used to estimate missing values ​​based on data from adjacent time points. In addition, there may be outliers in the data, such as data deviations caused by measurement errors or extreme weather events. Methods such as box plots can be used to identify and remove outliers. Box plots can intuitively display the distribution of data and identify data points that are beyond the normal range. In order to eliminate the differences in the numerical ranges between different factors, the data needs to be standardized. For example, the minimum-maximum standardization method can be used to scale the data of each factor to between 0 and 1 for better subsequent analysis. After data preprocessing is completed, it is necessary to determine the probability distribution type of each factor. For example, if you want to know the probability distribution of factors such as light intensity, wind speed, and temperature. Statistical test methods such as chi-square test and Kolmogorov-Smirnov test can be used to determine whether each factor obeys common probability distribution types such as normal distribution and Weibull distribution. The Kolmogorov-Smirnov test can compare the difference between the empirical distribution of sample data and the assumed theoretical distribution, thereby determining whether the sample data obeys the theoretical distribution. The hypothesis test results show that light intensity obeys normal distribution, wind speed obeys Weibull distribution, and temperature obeys gamma distribution. For factors that obey known distribution types, such as light intensity obeying normal distribution, the maximum likelihood estimation method can be used to estimate the corresponding distribution parameters. The basic idea of ​​the maximum likelihood estimation method is to find the parameter value that is most likely to produce the observed data. Specifically, a log-likelihood function is constructed based on the sample data, and the maximum likelihood estimates of the mean and variance are obtained by solving the log-likelihood equation group. Assume that the sample mean of light intensity calculated based on historical data is 800W / m 2 , the sample variance is 10000(W / m 2)2. For the wind speed that obeys the Weibull distribution, its scale parameter and shape parameter can be estimated. Assume that the scale parameter of the wind speed calculated according to historical data is 6m / s and the shape parameter is 2. For the temperature that obeys the gamma distribution, its shape parameter and scale parameter can be estimated. Assume that the shape parameter of the temperature calculated according to historical data is 3 and the scale parameter is 2. Substitute the estimated parameters into the corresponding probability density function expression to obtain the probability density function of each key factor. For example, the probability density function of light intensity is a normal distribution function, the probability density function of wind speed is a Weibull distribution function, and the probability density function of temperature is a gamma distribution function. After obtaining the probability density function, the characteristic quantities such as the mean and variance of the probability density function can be calculated by numerical integration and other methods. The mean reflects the average level of the factor, and the variance reflects the degree of fluctuation of the factor. In addition, the commonly used quantiles, such as P50, P90, etc., can be obtained by solving the inverse function of the probability distribution function. Quantiles can be used to describe the value range of factors at different probability levels. Using the estimated probability density function, a large number of random samples that obey the distribution can be generated by Monte Carlo simulation. Monte Carlo simulation is a numerical simulation method based on random sampling. Specifically, the probability distribution function is first obtained based on the probability density function, and then a random sample that obeys the target distribution is obtained by generating uniform random numbers on the interval [0, 1] and substituting them into the inverse function of the probability distribution function. For example, to generate 1000 random samples that obey the light intensity distribution, 1000 uniform random numbers on the interval [0, 1] can be generated first, and then these random numbers are substituted into the inverse function of the light intensity probability distribution function to obtain 1000 random samples of light intensity. Repeating the above process can obtain a sufficient number of random samples. Based on the generated random samples, the risk level and uncertainty of the system under different scenarios can be evaluated by statistical analysis and hypothesis testing. For example, suppose you want to evaluate the risk of renewable energy generation in the next 10 years. Through Monte Carlo simulation, 10,000 random samples of factors such as light intensity, wind speed, and temperature in the next 10 years are generated. Substituting these samples into the renewable energy power generation forecast model, we can get 10,000 forecast values ​​of renewable energy power generation in the next 10 years. By analyzing the distribution of these 10,000 forecast values, we can evaluate the risk level and uncertainty of renewable energy power generation. For example, we can calculate the mean and variance of the forecast values ​​to estimate the expected value and volatility of renewable energy power generation risk. We can calculate the quantiles of the forecast values ​​to estimate the risk threshold at a given confidence level. We can also use hypothesis testing to determine whether the difference in risk levels under different scenarios is significant. For factors that do not obey known distribution types, such as historical data of certain meteorological factors that do not obey any common probability distribution types, we can use non-parametric estimation and other methods to estimate their probability density functions directly based on sample data.For example, the kernel density estimation method can be used to fit the empirical distribution of the sample by selecting an appropriate kernel function and bandwidth parameter to obtain a non-parametric estimate of the probability density function. Kernel density estimation is a non-parametric method for estimating the probability density function based on sample data. It estimates the overall probability density function by taking a weighted average of the probability density around each sample point. The choice of kernel function and the setting of bandwidth parameters will affect the smoothness of the estimation result.

[0027] Step S103, using the uncertainty distribution function, a large number of input samples are generated by Monte Carlo random sampling method, and substituted into the potential output calculation model of renewable energy power generation to obtain the probability distribution of potential output. This probability distribution reflects the uncertainty characteristics of the potential output of renewable energy power generation.

[0028] According to the uncertainty distribution function, a large number of random input sample data that meet the distribution characteristics are generated by using the Monte Carlo random sampling algorithm. The generated random input sample data are preprocessed as necessary, including data cleaning, outlier processing and data normalization, and characteristic variables related to the potential output of renewable energy power generation are extracted. The preprocessed input sample data are substituted into the pre-established new energy power generation potential output calculation model, which can be a mathematical model based on physical mechanism or a data-driven model trained by machine learning algorithms (such as support vector machines or neural networks). The potential output value corresponding to each input sample is obtained through model calculation. According to the rated output and historical operation data of the renewable energy generator set, several potential output intervals are reasonably set. The calculated potential output value of each input sample is divided into the corresponding interval. The number of samples in each potential output interval is counted to obtain the frequency distribution of potential output in different value intervals. According to the frequency distribution of potential output, the probability of each potential output interval is calculated by using the empirical probability method to obtain the probability distribution of the potential output of renewable energy power generation. The probability distribution of the potential output of renewable energy power generation is analyzed, and the degree of uncertainty of the potential output is quantitatively evaluated through statistical indicators such as the skewness and kurtosis of the distribution curve. When the absolute value of the skewness is greater than 1 or the kurtosis is greater than 5, it indicates that there is a large uncertainty in the potential output of renewable energy power generation, and it is necessary to deal with its volatility risk through measures such as reasonable scheduling of spare capacity and optimization of energy storage configuration; when the absolute value of the skewness is less than 5 and the kurtosis is less than 3, it indicates that the uncertainty of the potential output of renewable energy power generation is small, and its grid connection ratio can be appropriately increased to reduce the phenomenon of wind and solar power abandonment.

[0029] For example, the core idea of ​​the Monte Carlo random sampling algorithm is to generate random samples from a known probability distribution. In the evaluation of renewable energy generation potential, the probability distribution functions of key influencing factors (such as light intensity, wind speed, and temperature) have been determined. Taking light intensity as an example, it is assumed to follow a normal distribution with a mean of 800W / m 2, standard deviation is 100W / m 2 . Using the Monte Carlo method, a large number of random light intensity samples can be generated, and the distribution of these sample values ​​conforms to the previously determined normal distribution. This is done to simulate the random variation characteristics of light intensity and make it closer to the actual situation. The generated sample data usually needs to be preprocessed to improve the accuracy and stability of the model. For example, the data can be cleaned to remove outliers and missing values. Suppose the light intensity sample value on a certain day is -100W / m 2 , which is obviously an outlier and needs to be removed or replaced. In addition, the data can be normalized to convert data of different dimensions into the same numerical range, such as converting light intensity, wind speed and temperature into values ​​between 0 and 1. This can prevent some features from having too much impact on the model due to their large values. Substituting the preprocessed sample data into the potential output calculation model of new energy power generation, the potential output value corresponding to each sample can be obtained. This model can be a mathematical model based on physical mechanism or a data-driven machine learning model. Assume that a simple linear regression model is adopted: potential output = a*light intensity + b*wind speed + c*temperature, where a, b, and c are model parameters. Substituting the light intensity, wind speed and temperature values ​​of each sample into the model, the corresponding potential output value can be calculated. In order to better analyze the distribution of potential output, it can be divided into different intervals. For example, the potential output can be divided into several intervals such as 0-10MW, 10-20MW, 20-30MW, etc. Then count the number of samples in each interval to obtain the frequency distribution of potential output. Doing so can intuitively understand the distribution of potential output in different numerical ranges. According to the frequency distribution, the probability of each potential output interval can be calculated. For example, assuming that among 1000 samples, 200 samples have potential outputs falling in the 10-20MW range, the probability of this interval is 200 / 1000=2. By calculating the probability of each interval, the probability distribution of the potential output of renewable energy power generation can be obtained. By analyzing statistical indicators such as the skewness and kurtosis of the probability distribution, the degree of uncertainty of the potential output can be quantitatively evaluated. Skewness measures the degree of skewness of the distribution, and kurtosis measures the degree of sharpness of the distribution. Assume that the calculated skewness of the potential output distribution is 8 and the kurtosis is 3. Since the absolute value of the skewness is less than 1 and the kurtosis is less than 5, it indicates that the uncertainty of the potential output is small. In this case, the proportion of renewable energy connected to the grid can be appropriately increased to reduce the phenomenon of wind and solar power abandonment. On the contrary, if the skewness or kurtosis is too large, it means that the uncertainty of the potential output is high, and measures such as energy storage and spare capacity are needed to deal with its volatility risk. For example, energy storage equipment can be configured to store excess electricity when power generation exceeds load demand, and release the stored energy when power generation is less than load demand, thereby smoothing the volatility of renewable energy power generation and improving the stability of the power grid.

[0030] Step S104, based on the uncertainty distribution function of the potential output of renewable energy power generation, a large number of random input sample data that meet the distribution characteristics are generated by using the Monte Carlo random sampling algorithm, and the potential output value corresponding to each input sample is calculated by substituting the preprocessed input sample data into a mathematical model based on physical mechanism or a data-driven model trained by a machine learning algorithm, thereby obtaining the probability distribution of the potential output of renewable energy power generation, and quantitatively judging the degree of uncertainty of the potential output based on statistical indicators such as the skewness and kurtosis of the distribution curve.

[0031] According to the uncertainty distribution function of the potential output of renewable energy power generation, the Monte Carlo random sampling algorithm is used to generate at least one thousand random input sample data that meet the distribution characteristics. The random input sample data generated by the Monte Carlo random sampling algorithm is preprocessed to remove outliers and missing values, and the maximum and minimum normalization method is used for normalization, and the data is mapped to the interval [0, 1] to obtain the preprocessed input sample data. According to the pre-established physical mechanism model of the potential output of renewable energy power generation, the gradient descent algorithm and the back propagation algorithm are used to use the preprocessed input sample data as training data, and the model parameters are optimized by minimizing the mean square error loss function to obtain the trained physical mechanism model. According to the pre-established data-driven model of the potential output of renewable energy power generation, the random forest regression algorithm is used to use the preprocessed input sample data as training data, and multiple decision trees are constructed through self-service sampling and random feature selection, and the prediction results of all decision trees are averaged to obtain the trained data-driven model. Substitute the preprocessed input sample data into the trained physical mechanism model and data-driven model respectively, calculate the potential output value corresponding to each input sample, and take the weighted average of the prediction results of the two models as the final potential output prediction value. According to the potential output prediction values ​​of all input samples, draw the probability distribution curve of the potential output of renewable energy power generation. Calculate the skewness coefficient and kurtosis coefficient of the distribution curve. When the absolute value of the skewness coefficient is greater than 0 and the kurtosis coefficient is greater than 0, it is judged that the potential output of renewable energy power generation has a high uncertainty; otherwise, the uncertainty of the potential output of renewable energy power generation is low.

[0032] Exemplarily, the Monte Carlo random sampling algorithm is an effective method to generate random sample data from a known probability distribution. In the field of renewable energy power generation, it can be used to simulate various uncertain factors that affect power generation output. For example, suppose that based on historical data and weather forecasts, the probability distribution function of light intensity in the next 24 hours is obtained, which may be a gamma distribution or a beta distribution. Then, the Monte Carlo method can be used to generate 1,000 random sample values ​​of light intensity. These sample values ​​conform to the predetermined probability distribution characteristics and can more realistically reflect the changes in light intensity. There may be outliers or missing values ​​in the generated sample data, which need to be preprocessed. Suppose that the light intensity sample value at a certain moment is -500W / m 2 , which is obviously an outlier and needs to be removed. If the light intensity data at a certain moment is missing, the data at adjacent moments can be used for interpolation or filling. In addition, in order to eliminate the impact of dimensions between different features, the data needs to be normalized. For example, normalize the light intensity, wind speed, temperature and other data to between 0 and 1. Assume that the light intensity at a certain moment is 600W / m 2 , the minimum value is 0 and the maximum value is 1000W / m 2, then the normalized value is (600-0) / (1000-0)=6. The calculation model of the potential output of new energy power generation can be a model based on physical mechanism or a data-driven model. The physical mechanism model usually establishes mathematical equations based on the principle of energy conversion and equipment characteristics. For example, the physical mechanism model of photovoltaic power generation can consider factors such as light intensity, solar panel area, and conversion efficiency. The data-driven model uses machine learning algorithms to learn the relationship between power generation output and various influencing factors from historical data. For example, a data-driven model can be constructed using the random forest regression algorithm. Suppose a physical mechanism model of the potential output of photovoltaic power generation is established, in which the model parameters include the area of ​​the solar panel, the conversion efficiency, etc. Using the gradient descent algorithm and the back propagation algorithm, these parameters can be optimized according to the preprocessed input sample data to minimize the error between the model's predicted value and the actual value. Suppose a data-driven model is constructed using the random forest regression algorithm. The algorithm constructs multiple decision trees through self-service sampling and random feature selection, and then averages the prediction results of all decision trees to obtain the final prediction result. For example, suppose 100 decision trees are constructed, and the prediction results of each tree are slightly different. Averaging these results can obtain more stable and accurate prediction values. Substituting the preprocessed input sample data into the trained physical mechanism model and data-driven model respectively, two prediction results can be obtained. In order to comprehensively utilize the advantages of the two models, the two prediction results can be weighted averaged. For example, assuming that the weight of the physical mechanism model is 6 and the weight of the data-driven model is 4, the final potential output prediction value is 6×physical mechanism model prediction value+4×data-driven model prediction value. According to the potential output prediction values ​​of all input samples, a probability distribution curve can be drawn. This curve can intuitively show the probability of potential output in different numerical ranges. For example, it can be observed that the probability of potential output is higher in a certain interval and lower in other intervals. By calculating the skewness coefficient and kurtosis coefficient of the probability distribution curve, the uncertainty of potential output can be quantitatively evaluated. The skewness coefficient measures the degree of skewness of the distribution, and the kurtosis coefficient measures the degree of sharpness of the distribution. For example, if the absolute value of the skewness coefficient is 8 and the kurtosis coefficient is 3, it means that the potential output distribution is slightly skewed and the peak is sharp, indicating high uncertainty. On the contrary, if the skewness coefficient is close to 0 and the kurtosis coefficient is close to 0, it means that the potential output distribution is relatively symmetrical and the peak is relatively gentle, indicating low uncertainty. According to the level of uncertainty, corresponding strategies can be formulated, such as configuring energy storage equipment or adjusting the grid connection ratio.

[0033] Step S105, perform wavelet analysis on the obtained probability distribution of potential output of renewable energy power generation, realize multi-scale decomposition, and obtain the output fluctuation characteristics at different time scales. When the fluctuation exceeds the preset threshold, it is judged as abnormal fluctuation and triggers an early warning. The preset threshold is determined based on historical data statistics and power system stability requirements. The abnormal fluctuation early warning results are used to guide subsequent power generation forecast corrections and energy storage strategy optimization.

[0034] Obtain the probability distribution data of the potential output of renewable energy power generation, and use non-parametric estimation methods such as kernel density estimation to obtain the probability density function. Discretize the probability density function and use it as the input signal of the wavelet transform. Perform wavelet packet decomposition on the input signal to obtain the wavelet coefficients at multiple time scales. According to the historical data of renewable energy power generation, analyze the statistical characteristics of the wavelet coefficients at each time scale, and combine the requirements of the power system stability index to determine the abnormal fluctuation threshold of the wavelet coefficients at each scale by using statistical hypothesis testing and other methods. Determine whether the wavelet coefficient at each time scale exceeds the corresponding abnormal fluctuation threshold. If it exceeds, mark the output fluctuation at this time scale as abnormal fluctuation. Summarize the abnormal fluctuation information at each time scale and construct the abnormal fluctuation feature vector. Use historical data, feature engineering and data preprocessing to construct the training sample of the power generation prediction model. Take the abnormal fluctuation feature vector as an additional input feature, use algorithms such as long short-term memory neural network to predict the renewable energy power generation, and obtain the revised power generation prediction result. According to the revised power generation forecast results, taking into account the real-time operation status and dispatch constraints of the power grid, with the minimum charging and discharging cost of the energy storage system as the optimization goal, heuristic algorithms such as particle swarm optimization are used to optimize the charging and discharging power and time of the energy storage system to obtain the optimal energy storage operation strategy. The optimized energy storage strategy is converted into control instructions for the energy storage system, which are sent to the energy storage device through the energy management system to achieve real-time dispatch control of the energy storage system. At the same time, according to the dispatch requirements of the power grid, the output of new energy power generation and the charging and discharging of the energy storage system are coordinated to improve the adaptability of the power grid to fluctuations in new energy power generation and ensure the safe and stable operation of the power system.

[0035] For example, the power prediction of renewable energy generation is the key to the stable operation of the power system. Since the output of renewable energy generation such as wind power and solar energy is intermittent and volatile, it is crucial to accurately predict its output and formulate corresponding energy storage strategies. Here we will introduce a method for renewable energy power generation power prediction and energy storage optimization based on wavelet packet decomposition and long short-term memory neural network. First, it is necessary to obtain the probability distribution data of the potential output of renewable energy power generation. For example, the potential output at each moment can be calculated by collecting wind speed data of wind farms over a period of time and combining it with the power curve of the wind turbine. Assume that 1000 hours of wind speed data are collected and 1000 potential output data are calculated. Using the kernel density estimation method, the probability density function of the potential output can be obtained. The kernel density estimation method can be understood as fitting the distribution of sample data with small normal distribution curves, and finally obtaining a smooth curve, which is the probability density function. Discretize the probability density function, for example, divide the output range into 10 intervals, each interval corresponds to a probability value, and obtain a discretized probability distribution as the input signal of the wavelet transform. Next, the discretized probability distribution is decomposed by wavelet packets. Wavelet packet decomposition can decompose the signal into components at different time scales, such as decomposing the probability distribution into high-frequency, medium-frequency and low-frequency parts. The high-frequency part reflects the fluctuation of rapid output change, and the low-frequency part reflects the trend of slow output change. Assume that the probability distribution is decomposed into four time scales through wavelet packet decomposition, representing the fluctuation periods of 1 hour, 4 hours, 12 hours and 24 hours respectively. Then, it is necessary to determine the abnormal fluctuation threshold of the wavelet coefficient at each time scale. For example, the mean and standard deviation of the wavelet coefficient at each time scale can be counted based on historical data. Assuming that the mean of the wavelet coefficient at the 1-hour scale is 0 and the standard deviation is 1, the abnormal fluctuation threshold can be set to 2, that is, the fluctuation exceeding 2 standard deviations from the mean is considered to be an abnormal fluctuation. Similarly, the abnormal fluctuation threshold at other time scales can be determined. Next, it is determined whether the wavelet coefficient at each time scale exceeds the corresponding abnormal fluctuation threshold. If it exceeds, the output fluctuation at that time scale is marked as abnormal fluctuation. For example, if the wavelet coefficient at a certain moment at the 1-hour scale is 3, which exceeds the threshold 2, it is marked as an abnormal fluctuation. Abnormal fluctuation information of all time scales is aggregated to construct abnormal fluctuation feature vectors. For example, the abnormal fluctuations of four time scales can be represented by a four-dimensional vector, where 1 represents abnormal fluctuations and 0 represents normal fluctuations. Use historical data to construct training samples for the power generation prediction model. For example, collect power generation data and weather data from the past year, and perform data preprocessing, such as removing outliers and filling missing values. Use the abnormal fluctuation feature vector as an additional input feature, and use algorithms such as long short-term memory neural networks for power prediction. Long short-term memory neural networks are suitable for processing time series data and can capture long-term dependencies in the data.According to the revised power generation forecast results, the real-time operation status of the power grid and the scheduling constraints, such as the maximum charging and discharging power and energy storage capacity limit, are considered. The charging and discharging power and time of the energy storage system are optimized by using heuristic algorithms such as particle swarm optimization, with the minimum charging and discharging cost of the energy storage system as the optimization goal. For example, when it is predicted that the power generation will be high in the future, the energy storage system can be controlled to charge and store the excess energy; when it is predicted that the power generation will be low, the energy storage system can be controlled to discharge and supplement the power shortage. Finally, the optimized energy storage strategy is converted into the control instructions of the energy storage system, which are sent to the energy storage device through the energy management system to realize the real-time scheduling control of the energy storage system. For example, the information such as charging power, discharging power, charging time, and discharging time is converted into control signals and sent to the energy storage converter and other equipment. By coordinating the output of new energy power generation and the charging and discharging of the energy storage system, the volatility of new energy power generation can be effectively smoothed, the adaptability of the power grid to the fluctuation of new energy power generation can be improved, and the safe and stable operation of the power system can be guaranteed.

[0036] Step S106, obtain the probability density function through kernel density estimation, discretize it as the wavelet transform input signal, use wavelet packet decomposition to obtain multi-time scale wavelet coefficients, determine the abnormal fluctuation threshold of each scale according to historical data, judge whether the wavelet coefficient exceeds the threshold to obtain abnormal fluctuation, and summarize the abnormal fluctuations to construct a feature vector.

[0037] The probability density function of the data is obtained by kernel density estimation, and then the continuous probability density function is discretized using the histogram method to obtain a discretized probability density function for subsequent wavelet transform processing. The discretized probability density function is used as the input signal of the wavelet transform, and the Daubechies wavelet is used to decompose the input signal at multiple scales to obtain wavelet coefficients at multiple time scales. According to historical data, the abnormal fluctuation threshold is calculated for the wavelet coefficients at each time scale. The specific method is: take the absolute value of the wavelet coefficients of the historical data, and then calculate its mean and standard deviation. The abnormal fluctuation threshold is set to the mean plus 3 times the standard deviation. If the wavelet coefficient at the current moment exceeds the abnormal fluctuation threshold at this scale, it is judged that there is an abnormal fluctuation at this moment. The abnormal fluctuations detected at each time scale are summarized, and the statistical characteristics of the abnormal fluctuations are extracted, including the time scale, duration, amplitude, etc. of the abnormal fluctuations, and the abnormal fluctuation feature vector is constructed. The random forest algorithm is used to classify the abnormal fluctuation feature vector, and the anomaly detection model is trained to determine whether the new input data is abnormal. If the abnormal fluctuation feature vector of the new input data is judged as abnormal by the anomaly detection model, an early warning is triggered, indicating the risk of abnormal fluctuations. According to the time scale and amplitude of abnormal fluctuations, abnormal fluctuations are graded and warned. The larger the time scale and the larger the amplitude of abnormal fluctuations, the higher the risk level. According to the preset risk level threshold, abnormal fluctuations are divided into different levels such as low risk, medium risk, and high risk, indicating different levels of risks and providing a basis for business decision-making.

[0038] Exemplarily, kernel density estimation is a non-parametric method for estimating probability density functions. It can be understood as fitting the distribution of sample data with small normal distribution curves (kernel functions), and finally superimposing these small normal distribution curves to obtain a smooth curve, which is the probability density function. For example, 1000 hours of wind speed data of a wind farm were collected, and 1000 potential output data were calculated based on the power curve of the wind turbine. Using the kernel density estimation method, the probability density function of the potential output of the wind farm can be obtained, which describes the probability of occurrence of different output levels. After obtaining the probability density function, it needs to be discretized for subsequent wavelet transform processing. The histogram method is a commonly used discretization method. It divides the value range of the continuous probability density function into several intervals, counts the frequency of data occurrence in each interval, and converts the frequency into a probability value, and finally obtains a discretized probability density function. For example, the range of potential output of a wind farm can be divided into 10 intervals, representing output levels of 0-10%, 10%-20%...90%-100%. Then, the frequency of occurrence of potential output data in each interval is counted and converted into probability values, and finally a discretized probability density function containing 10 probability values ​​is obtained. The discretized probability density function is used as the input signal of the wavelet transform, and the Daubechies wavelet is used for multi-scale decomposition. Daubechies wavelet is a commonly used wavelet basis function with good time-frequency localization characteristics. Multi-scale decomposition can decompose the signal into components at different time scales, such as high-frequency, medium-frequency and low-frequency parts. The high-frequency part reflects the fluctuation of rapid changes in output, and the low-frequency part reflects the trend of slow changes in output. For example, the discretized probability density function can be decomposed into 4 time scales, representing the fluctuation periods of 1 hour, 4 hours, 12 hours and 24 hours respectively. In order to determine whether there is abnormal fluctuation, it is necessary to determine the abnormal fluctuation threshold of the wavelet coefficient at each time scale. A commonly used method is to calculate the mean and standard deviation based on the wavelet coefficients of historical data, and set the value of the mean plus 3 times the standard deviation as the abnormal fluctuation threshold. For example, the wavelet coefficient data of the past year is collected, the absolute value of the wavelet coefficient at the 1-hour time scale is taken, and its mean and standard deviation are calculated. Assuming the mean is 5 and the standard deviation is 1, the abnormal fluctuation threshold is 5+3*1=8. If the wavelet coefficient at the current moment exceeds the abnormal fluctuation threshold at this scale, it is judged that there is an abnormal fluctuation at this moment. For example, the wavelet coefficient at the 1-hour time scale at the current moment is 9, which exceeds the threshold 8, so it is judged that there is an abnormal fluctuation at this moment. Similarly, it can be judged whether there is an abnormal fluctuation at other time scales. Summarize the abnormal fluctuations detected at each time scale, extract the statistical characteristics of the abnormal fluctuations, and construct the abnormal fluctuation feature vector. The abnormal fluctuation feature vector can include information such as the time scale, duration, and amplitude of the abnormal fluctuation.For example, if abnormal fluctuations are detected at both 1-hour and 4-hour time scales, with durations of 2 hours and 4 hours respectively and amplitudes of 1 and 2 respectively, an abnormal fluctuation feature vector containing this information can be constructed. The abnormal fluctuation feature vector is classified using the random forest algorithm to train the anomaly detection model. Random forest is an ensemble learning method that improves the accuracy of classification by constructing multiple decision trees and voting. For example, a random forest anomaly detection model can be trained using data from the past year, with the abnormal fluctuation feature vector as input and the presence or absence of abnormal fluctuations as output. When the abnormal fluctuation feature vector of the new input data is judged as abnormal by the anomaly detection model, an early warning is triggered. According to the time scale and amplitude of the abnormal fluctuation, a graded early warning is given to the abnormal fluctuation. The larger the time scale and the larger the amplitude of the abnormal fluctuation, the higher the risk level. For example, an abnormal fluctuation with an amplitude of 1 at a 1-hour time scale may be classified as low risk, while an abnormal fluctuation with an amplitude of 5 at a 24-hour time scale may be classified as high risk. According to the preset risk level threshold, the abnormal fluctuation is classified into different levels such as low risk, medium risk, and high risk, prompting different levels of risk and providing a basis for business decision-making. Doing so can help better identify and respond to abnormal fluctuations in renewable energy power generation output and ensure the safe and stable operation of the power system.

[0039] Step S107, for photovoltaic power generation, a nonlinear mapping relationship between factors such as solar radiation intensity, photovoltaic module conversion efficiency, temperature coefficient, aging attenuation rate and power generation efficiency is established using a deep learning algorithm to dynamically correct the predicted value of photovoltaic power generation potential output. The correction process takes into account the abnormal fluctuation warning results to improve the prediction accuracy.

[0040] The historical and real-time data of multiple factors affecting photovoltaic power generation efficiency, including solar radiation intensity, photovoltaic module conversion efficiency, temperature coefficient and aging attenuation rate, were obtained. The historical data obtained were preprocessed using Python's Pandas library to remove missing values ​​and outliers, and the data were normalized to a unified scale of 0-1 using the minimum-maximum normalization method. The long short-term memory (LSTM) neural network model was built using Python's Keras library, with the historical data of the influencing factors as input and the actual power generation efficiency of the corresponding time period as output, and the nonlinear mapping relationship between the influencing factors and the power generation efficiency was established through training. The ARIMA time series model was used to predict the values ​​of solar radiation intensity, photovoltaic module conversion efficiency, temperature coefficient and aging attenuation rate in the future based on the historical data of each influencing factor. The predicted influencing factor data was input into the trained LSTM model to obtain the predicted value of photovoltaic power generation potential output in the time period. The isolation forest algorithm in Python's Scikit-learn library was used to perform anomaly detection with the real-time data of the influencing factors as input. The normal fluctuation range of each influencing factor was calculated based on the historical data. If the real-time data exceeds this range, it is judged as an abnormal fluctuation and an early warning information is generated. Conditionally merge the abnormal fluctuation warning information with the photovoltaic power generation potential output forecast value generated by the LSTM model. If the data of a certain influencing factor is judged to be abnormal, the power generation output forecast value of the corresponding time is adjusted downward by a fixed proportion in the prediction result of the LSTM model to obtain the revised forecast value. Using Python's Sklearn library, the revised photovoltaic power generation output forecast value is compared with the actual value to calculate the mean absolute percentage error (MAPE). If the MAPE is greater than 10%, return to step 3 and retrain the LSTM model using the updated historical data; otherwise, output the final photovoltaic power generation output forecast result.

[0041] For example, photovoltaic power generation efficiency is affected by many factors, including solar radiation intensity, photovoltaic module conversion efficiency, temperature coefficient and aging attenuation rate. In order to accurately predict photovoltaic power generation output, it is necessary to obtain historical data and real-time data of these influencing factors. Historical data can be obtained through meteorological stations, photovoltaic power station monitoring systems and other channels, and real-time data can be obtained through real-time measurement by sensors. For example, data such as solar radiation intensity and photovoltaic module temperature can be recorded every 15 minutes. After obtaining the data, the historical data needs to be preprocessed. The Pandas library can be used to easily clean and convert data. For example, the fillna function of Pandas can be used to fill missing values, and the drop function can be used to remove outliers. In addition, in order to eliminate the impact of dimensions between different data, the data needs to be normalized, such as using the minimum-maximum normalization method to scale the data to between 0 and 1. Assuming that the highest temperature on a certain day is 35 degrees Celsius and the lowest temperature is 20 degrees Celsius, then 25 degrees Celsius is (25-20) / (35-20)=33 after normalization. Using the preprocessed historical data, a long short-term memory neural network model can be constructed to predict photovoltaic power generation efficiency. The LSTM neural network is a special recurrent neural network that can learn long-term dependencies in time series data. The historical data of influencing factors is used as input, and the actual power generation efficiency of the corresponding time period is used as output. The LSTM model is trained to establish a nonlinear mapping relationship between influencing factors and power generation efficiency. For example, the model can be trained with data from the past year, with the input being the daily solar radiation intensity, temperature and other data, and the output being the daily photovoltaic power generation efficiency. In order to predict the photovoltaic power generation output in the future, it is necessary to first predict the future values ​​of the influencing factors. The ARIMA time series model can be used to predict factors such as solar radiation intensity and temperature. The ARIMA model can capture the autocorrelation in time series data and make predictions based on it. For example, the solar radiation intensity in the next week can be used to predict the solar radiation intensity in the past month. By inputting the influencing factor data predicted by the ARIMA model into the trained LSTM model, the predicted value of the potential photovoltaic power generation output in the future can be obtained. For example, by inputting the predicted solar radiation intensity, temperature and other data in the next week into the LSTM model, the predicted value of the photovoltaic power generation output in the next week can be obtained. In order to improve the accuracy of the prediction, it is necessary to detect anomalies in the real-time data of the influencing factors. The Isolation Forest Algorithm can be used to identify abnormal data. The Isolation Forest Algorithm is an algorithm based on anomaly isolation, which can effectively identify data points that deviate from the normal range. For example, if the solar radiation intensity at a certain moment suddenly drops to a very low value, the Isolation Forest Algorithm can identify it as abnormal data. Based on historical data, the normal fluctuation range of each influencing factor can be calculated.For example, the mean and standard deviation of solar radiation intensity in historical data can be calculated, and the range of the mean plus or minus two times the standard deviation is taken as the normal fluctuation range. If the real-time data exceeds this range, it is judged as abnormal fluctuation and an early warning message is generated. The abnormal fluctuation early warning information is conditionally merged with the prediction results of the LSTM model. If a certain influencing factor is judged to be abnormal, the power generation output forecast value of the corresponding time is adjusted downward by a fixed proportion in the prediction results of the LSTM model. This is done to avoid abnormal data from having too much impact on the prediction results. For example, if the real-time solar radiation intensity is judged to be abnormally low, the photovoltaic power generation output forecast value of the corresponding time period is adjusted downward by 10%. Finally, the Sklearn library is used to calculate the mean absolute percentage error (MAPE) between the corrected photovoltaic power generation output forecast value and the actual value. MAPE is a common indicator for measuring prediction accuracy. If MAPE is greater than 10%, it means that the prediction accuracy of the model is not high enough, and the LSTM model needs to be retrained using updated historical data. Otherwise, the final photovoltaic power generation output prediction result is output. This can continuously improve the prediction performance of the model and improve the accuracy of the prediction.

[0042] Step S108: For wind power generation, a computational fluid dynamics model is used to simulate the wind field distribution, obtain the wind speed and wind direction parameters at the location of the wind turbine, and determine the potential wind power output in combination with the wind turbine power curve. The abnormal fluctuation warning results are also considered for correction to improve the accuracy of wind power output prediction.

[0043] According to the three-dimensional model of the wind farm, the speed and direction distribution of the wind farm is simulated by the computational fluid dynamics method to obtain the wind speed and wind direction data at the location of the wind turbine. The obtained wind speed data is input into the power curve of the wind turbine to calculate the theoretical power generation of the wind turbine at this wind speed as the preliminary estimate of the potential wind power output. The support vector machine (SVM) algorithm is used to train the historical wind speed and wind direction data and the corresponding actual power generation data to establish a wind power output prediction model. The mean and standard deviation of the current wind speed and wind direction and the historical data are calculated to determine whether they exceed the normal range (such as 3 times the standard deviation). If they exceed, an early warning is triggered. According to the early warning level, the wind power output prediction value is corrected accordingly, such as reducing the weight coefficient of the prediction value. The corrected wind power output prediction value and the preliminary estimate are weighted averaged, and the weight can be determined according to the confidence of the two to obtain the final wind power output prediction result. The cross-validation method is used to evaluate the wind power output prediction model, and evaluation indicators such as the root mean square error (RMSE) are calculated. The prediction model is continuously optimized by adjusting parameters such as the kernel function type and penalty coefficient of the SVM. The optimized wind power output prediction model is applied to the actual wind farm. Combined with the real-time monitoring data of the wind turbine (such as wind speed, wind direction, generator speed, etc.), the prediction results are dynamically adjusted to provide a decision-making basis for the power generation plan formulation and grid-connected scheduling of the wind farm.

[0044] For example, wind power output prediction is crucial for the effective operation of wind farms. In order to more accurately predict the output of wind farms, a more accurate prediction model can be established based on the three-dimensional model of the wind farm, combined with computational fluid dynamics (CFD) methods and machine learning algorithms. First, using the three-dimensional model of the wind farm, the computational fluid dynamics method is used to simulate the wind speed and wind direction distribution of the wind farm. For example, professional CFD software such as ANSYS Fluent or OpenFOAM can be used to model and simulate the wind farm, input terrain data, surface roughness and other parameters, and simulate the wind speed and wind direction at various locations in the wind farm under different wind speeds and wind directions. Assume that the simulation results show that at a certain moment, the wind speed at the location of wind turbine A is 8 meters per second and the wind direction is due south. Then, the acquired wind speed data is input into the power curve of the wind turbine to calculate the theoretical power generation of the wind turbine at this wind speed. Each wind turbine has its own unique power curve, which describes the relationship between wind speed and power generation. For example, the power curve of wind turbine A shows that when the wind speed is 8 meters per second, its theoretical power generation is 5 megawatts. This value can be used as a preliminary estimate of the potential output of wind power. In order to improve the prediction accuracy, machine learning algorithms, such as support vector machines (SVM), can be used to train historical wind speed, wind direction data and corresponding actual power generation data to establish a wind power output prediction model. For example, collect wind speed, wind direction and actual power generation data for the past year, input these data into the SVM model for training, and learn the complex nonlinear relationship between wind speed, wind direction and actual power generation. When making a prediction, calculate the mean and standard deviation of the current wind speed and wind direction and historical data to determine whether they exceed the normal range. For example, the mean of the wind speed in the past year is 6 meters per second, and the standard deviation is 2 meters per second. If the current wind speed is 12 meters per second, which exceeds the mean plus 3 times the standard deviation, an early warning is triggered. The early warning level can be set according to the multiple of the excess, for example, exceeding 3 times the standard deviation is a first-level early warning, and exceeding 4 times the standard deviation is a second-level early warning. According to the early warning level, the wind power output forecast value is corrected accordingly. For example, if a first-level early warning is triggered, the weight coefficient of the predicted value is reduced, such as reducing the weight coefficient from 8 to 6. This is because when the wind speed exceeds the normal range, the actual power generation of the wind turbine may be affected by other factors. For example, excessively high wind speed may cause the wind turbine to operate at limited power. The output value predicted by the SVM model is weighted averaged with the preliminary estimate calculated based on the power curve to obtain the final wind power output prediction result. For example, the output value predicted by the SVM model is 4 MW, and the preliminary estimate is 5 MW. Assuming that the weight coefficients of the two are 6 and 4 respectively, the final prediction result is (4×6)+(5×4)=44 MW. The weight coefficient can be determined based on the confidence of the two. For example, if historical data shows that the SVM model has a higher prediction accuracy, it is given a higher weight. In order to evaluate and optimize the prediction model, the cross-validation method is used to calculate evaluation indicators such as the root mean square error (RMSE).By adjusting the SVM kernel function type, penalty coefficient and other parameters, the prediction model is continuously optimized. For example, different kernel functions, such as linear kernel, polynomial kernel, and radial basis kernel function, can be tried, and the value of the penalty coefficient C can be adjusted to compare the RMSE under different parameter combinations and select the parameter combination with the smallest RMSE. Finally, the optimized wind power output prediction model is applied to the actual wind farm. Combined with the real-time monitoring data of the wind turbine, such as wind speed, wind direction, generator speed, etc., the prediction results are dynamically adjusted to provide a decision-making basis for the wind farm's power generation plan formulation and grid-connected scheduling. For example, if the real-time monitoring data shows that a wind turbine has failed, the output prediction value of the wind turbine needs to be adjusted according to the actual situation.

[0045] Step S109, based on the potential output prediction results of photovoltaic and wind power generation, the energy storage capacity configuration and operation strategy are solved through the intelligent optimization algorithm, and the benefits of the energy storage system are maximized under the constraints of new energy consumption and system stability, and the dynamic prediction value of the energy storage consumption capacity index is obtained. The energy storage consumption capacity index is defined as the ratio of the new energy power generation that can be consumed by the energy storage system per unit time to the total power generation, which is used to quantify the contribution of the energy storage system to the new energy consumption.

[0046] According to the historical data and weather forecast data of photovoltaic power generation and wind power generation, data cleaning and preprocessing are first performed to remove outliers and missing values, and feature selection and feature engineering are performed to construct a suitable feature set. Then, the support vector machine algorithm is used to establish a potential output prediction model for photovoltaic and wind power generation. The model hyperparameters are optimized through cross-validation and grid search methods to improve the prediction accuracy of the model and obtain the potential output prediction value in the future. According to the potential output prediction value and the grid load demand, the particle swarm optimization algorithm is used to solve the optimal energy storage capacity configuration and operation strategy under the conditions of energy storage capacity constraints, new energy consumption constraints and system stability constraints. The objective function is to maximize the benefits of the energy storage system. The particle swarm algorithm searches for the optimal solution through iterative optimization to obtain the energy storage capacity configuration and operation strategy. According to the energy storage capacity configuration and operation strategy obtained by optimization, the charging and discharging power of the energy storage system in different periods is obtained through simulation calculation, and then the new energy power generation that can be absorbed by the energy storage system in unit time is obtained. The total power generation data of photovoltaic and wind power generation in the same period is obtained from the power grid dispatching system, and the ratio of the renewable energy power generation that can be absorbed by the energy storage system to the total power generation is calculated to obtain the energy storage absorption capacity index. The long short-term memory network model is used to dynamically predict the energy storage absorption capacity index. The model is trained with historical data, and the sliding window method is used to perform online updates and predictions to obtain the predicted value for a period of time in the future. According to the preset threshold, it is judged whether the predicted value of the energy storage absorption capacity index exceeds the threshold. If it exceeds, it indicates that the energy storage system contributes more to the absorption of renewable energy, otherwise it indicates that the contribution is smaller. According to the judgment result, the model predictive control method is used to dynamically optimize and adjust the energy storage capacity configuration and operation strategy. Through the rolling optimization method, the optimization objectives and constraints are updated in real time, and a new optimal control strategy is generated. It is applied to the actual operation control of the energy storage system to adapt to the changes in renewable energy output and load demand, continuously improve the contribution of the energy storage system to the absorption of renewable energy, and ensure the safe and stable operation of the power grid.

[0047] For example, data cleaning and preprocessing are the basis of the prediction model. Taking wind power generation data as an example, the original data may contain information such as wind speed, wind direction, temperature, humidity, air pressure, and the corresponding power generation. First, it is necessary to remove abnormal values, such as the instantaneous wind speed suddenly changes to zero or negative due to sensor failure. At the same time, to deal with missing values, such as missing data caused by data acquisition system failure in certain time periods, interpolation or mean method can be used to fill. Feature selection and feature engineering are to build a more effective feature set. For example, wind speed and wind direction have a greater impact on wind power generation, and these two features can be selected as the main features. In addition, new features can be constructed, such as the daily variation of wind speed and the stability of wind direction, to improve the prediction accuracy of the model. The support vector machine algorithm can be used to establish a potential output prediction model for photovoltaic and wind power generation. Taking photovoltaic power generation as an example, historical light intensity, temperature and other data can be used as input features, and actual power generation can be used as the output target to train a support vector machine regression model. Through methods such as cross-validation and grid search, the model's hyperparameters, such as kernel function type, penalty coefficient, etc., can be optimized to improve the model's prediction accuracy. Assume that the model predicts that the potential photovoltaic power generation in the next week is 100 MW, 120 MW, 110 MW, 90 MW, 80 MW, 100 MW, and 130 MW respectively. Particle swarm optimization algorithm can be used to solve the optimal energy storage capacity configuration and operation strategy. Taking a scenario including a wind farm and an energy storage system as an example, the objective function can be set to maximize the benefits of the energy storage system, that is, maximize the amount of new energy consumption and reduce the amount of wind / light abandonment. Constraints can include energy storage capacity constraints, such as the maximum charging and discharging rates of the energy storage system, and system stability constraints, such as the range of grid frequency fluctuations. Assuming that the optimal energy storage capacity obtained by the particle swarm optimization algorithm is 50 MWh, the best operation strategy is to store energy in the energy storage system when wind power generation is sufficient during the day, and release energy to the grid during the peak power consumption period at night. The energy storage absorption capacity index is an important indicator for evaluating the contribution of energy storage systems to the absorption of new energy. Assuming that the total amount of wind power generation in a certain period is 200 MW, and the wind power generation that the energy storage system can absorb is 80 MW, the energy storage absorption capacity index is 40%. Through the long short-term memory network model, the energy storage absorption capacity index can be dynamically predicted. Assume that the energy storage absorption capacity index for the next week is predicted to be 30%, 40%, 35%, 25%, 20%, 30%, and 45% respectively. The model predictive control method can be used to dynamically optimize the energy storage capacity configuration and operation strategy. Assume that the preset energy storage absorption capacity index threshold is 30%. According to the prediction results, the energy storage absorption capacity index of the first, second, third and last day of the next week exceeds the threshold, indicating that the energy storage system contributes more to the absorption of new energy. The energy storage absorption capacity index of the fourth and fifth days is lower than the threshold, indicating that the energy storage system contributes less to the absorption of new energy.At this time, the model predictive control method can be used to adjust the operation strategy of the energy storage system, such as increasing the charging rate of the energy storage system, or adjusting the dispatching strategy of the power grid to increase the absorption of new energy.

[0048] Step S1010, the prediction results of multiple models are combined by using an adaptive weighted fusion strategy, and the weight coefficients of each sub-model are dynamically adjusted according to the prediction error to improve the overall accuracy and robustness of the prediction of the energy storage capacity index. The prediction results of multiple models include the prediction results of statistical models based on historical data, dynamic models based on physical mechanisms, and data-driven models based on machine learning.

[0049] According to the historical data of the energy storage capacity index, an ARIMA time series model is established. By fitting the historical data, the trend of the energy storage capacity index in the future is predicted. Using the laws of physics and the operating mechanism of the power system, a differential equation group describing the operating state of the energy storage system is established. The equation group is solved to obtain the predicted value of the physical mechanism model of the energy storage capacity index. Based on the key factors affecting the energy storage capacity index, such as power load and renewable energy power generation, a BP neural network model is constructed. By training the network, a nonlinear mapping relationship between the key factors and the energy storage capacity index is established to achieve data-driven prediction. The root mean square error and mean absolute percentage error are used to quantitatively evaluate the prediction error of each sub-model. According to the size of the prediction error, the weighted average method is used to determine the initial weight coefficient of each sub-model. The smaller the error, the larger the weight coefficient. In the new time period, the latest prediction values ​​of the energy storage capacity index by the ARIMA model, physical mechanism model and BP neural network model are obtained again. The root mean square error and mean absolute percentage error of each sub-model in the latest period of time are calculated. According to the change of prediction error, the adaptive weighted fusion strategy is used to dynamically adjust the weight coefficient of each model. Specifically, the weight coefficient of the model with reduced prediction error increases proportionally; the weight coefficient of the model with increased prediction error decreases proportionally. Using the adjusted weight coefficient, the prediction results of the ARIMA model, the physical mechanism model and the BP neural network model are weighted and fused to obtain the final prediction value of the energy storage capacity index. The prediction error in the current time period is calculated to determine whether it exceeds the preset threshold. If it exceeds the threshold, it is considered that the current weight coefficient is no longer applicable and needs to be readjusted, and the return is made to step 6; if it does not exceed the threshold, it is considered that the current weight coefficient is still valid, and the existing weight coefficient is maintained, and the return is made to step 5 to continue the prediction of the next time period. Repeat steps 5-8 to achieve adaptive dynamic prediction of the energy storage capacity index. Through continuous error evaluation and weight adjustment, the prediction accuracy is continuously improved, and the adaptability of the prediction model to the changing trend of the energy storage capacity index is enhanced.

[0050] For example, the ARIMA model is a commonly used time series prediction method. It analyzes the autocorrelation and partial autocorrelation of time series data and establishes a mathematical model to describe the change law of data. For example, the energy storage capacity index data for each day in the past year can be collected, and the ARIMA model can be used to fit these historical data to predict the energy storage capacity index for the next week. Assume that according to the historical data, the ARIMA model selected is ARIMA (1, 1, 1), which means that the model contains an autoregressive term, a first-order difference term, and a moving average term. By fitting the historical data, the parameters of the model can be obtained, and these parameters can be used to predict the energy storage capacity index for the next week. The physical mechanism model is established based on the physical laws of the operation of the energy storage system and the operation mechanism of the power system. It usually uses a group of differential equations to describe the charging and discharging process of the energy storage system, the energy conversion process, and the interaction with the power grid. For example, a power system model containing an energy storage system, a photovoltaic power station, and a wind power station can be established. The model can take into account the volatility of photovoltaic and wind power generation, the change of power load, and the charging and discharging characteristics of the energy storage system. By solving the differential equations, the operating status of the energy storage system can be obtained, and then the energy storage absorption capacity index can be calculated. Assume that in a specific scenario, the photovoltaic power generation is 100 MW, the wind power generation is 80 MW, the power load is 150 MW, the energy storage system capacity is 50 MWh, and the initial state is 25 MWh. By solving the differential equations, the state of the energy storage system after one hour can be obtained, for example, the remaining power of the energy storage system is 35 MWh. Based on this information, the energy storage absorption capacity index within the hour can be calculated. The BP neural network model is a data-driven method that can establish a nonlinear mapping relationship between input variables and output variables. In this example, power load, new energy power generation, energy storage system capacity, etc. can be selected as input variables, and the energy storage absorption capacity index can be selected as the output variable. Collect historical data, such as daily data in the past year, to train the BP neural network model. Through training, the network can learn the complex relationship between input variables and output variables. For example, when the power load is high and the power generation of renewable energy is also high, the energy storage capacity index may be high; when the power load is low and the power generation of renewable energy is also low, the energy storage capacity index may be low. The trained BP neural network model can be used to predict the future energy storage capacity index. Root mean square error and mean absolute percentage error are commonly used prediction error evaluation indicators. The root mean square error reflects the degree of deviation between the predicted value and the actual value, and the mean absolute percentage error reflects the relative degree of deviation between the predicted value and the actual value. For example, the ARIMA model predicts that the energy storage capacity index for the next week will be 30%, 40%, 35%, 25%, 20%, 30%, and 45% respectively. Assume that the actual values ​​are 32%, 38%, 36%, 27%, 22%, 28%, and 42%.The root mean square error and mean absolute percentage error of the ARIMA model can be calculated separately and compared with the errors of other models. According to the size of the prediction error, the initial weight coefficient of each sub-model can be determined. The smaller the error, the larger the weight coefficient. For example, assuming that the root mean square error of the ARIMA model is the smallest, it is given the largest weight coefficient; the root mean square error of the BP neural network model is the largest, so it is given the smallest weight coefficient. In the new time period, the latest prediction values ​​of each sub-model are re-obtained, and the new prediction error is calculated. According to the change of the prediction error, the weight coefficient of each model is dynamically adjusted. For example, if the prediction error of the ARIMA model decreases, its weight coefficient is increased; if the prediction error of the BP neural network model increases, its weight coefficient is reduced. Using the adjusted weight coefficient, the prediction results of each sub-model are weighted and fused to obtain the final prediction value of the energy storage capacity index. For example, assuming that the prediction value of the ARIMA model is 35%, the weight coefficient is 5; the prediction value of the physical mechanism model is 40%, the weight coefficient is 3; the prediction value of the BP neural network model is 30%, and the weight coefficient is 2. The final prediction value is 35%*5+40%*3+30%*2=35%. Through continuous error evaluation and weight adjustment, the prediction accuracy can be continuously improved, and the prediction model's ability to adapt to the changing trend of the energy storage capacity index can be enhanced. Doing so can effectively utilize the advantages of different models, improve the accuracy and reliability of the overall prediction, and provide a more reliable decision-making basis for the optimized operation of the energy storage system.

[0051] The above is only a specific implementation method of this specification. Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working process of the system, module and unit described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here. It should be understood that the protection scope of this specification is not limited to this. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed in this specification, and these modifications or replacements should be covered within the protection scope of this specification.

Claims

1. A method for predicting the potential output of renewable energy power generation, characterized in that: include: Acquire historical meteorological data and geographic information data, and based on the historical meteorological data and geographic information data, establish a multivariate regression model of key factors and weight coefficients affecting the potential output of renewable energy power generation, use the Monte Carlo random sampling method to generate a large number of input samples, substitute the input samples into the renewable energy power generation potential output calculation model to obtain the probability distribution of potential output, perform wavelet analysis on the probability distribution to obtain output fluctuation characteristics at different time scales, and perform abnormal fluctuation warning according to the output fluctuation characteristics; For photovoltaic power generation, a nonlinear mapping relationship between solar radiation intensity, photovoltaic module conversion efficiency, temperature coefficient, aging attenuation factor and power generation efficiency is established to dynamically correct the predicted value of photovoltaic power generation potential output; For wind power generation, a computational fluid dynamics model is used to simulate the wind field distribution, obtain the wind speed and wind direction parameters at the location of the wind turbine, determine the potential output of wind power in combination with the wind turbine power curve, and make corrections based on the abnormal fluctuation warning results; Based on the potential output forecast results of photovoltaic and wind power generation, the energy storage capacity configuration and operation strategy are solved through intelligent optimization algorithms to maximize the benefits of the energy storage system, and the dynamic forecast value of the energy storage absorption capacity index is obtained; And an adaptive weighted fusion strategy is used to combine the prediction results of statistical models based on historical data, dynamic models based on physical mechanisms, and data-driven models based on machine learning to predict the energy storage absorption capacity index.

2. The method for predicting the potential output of renewable energy power generation according to claim 1, characterized in that: The Monte Carlo random sampling method is used to generate a large number of input samples, including: According to the uncertainty distribution function of the key factor, a Monte Carlo random sampling algorithm is used to generate a large amount of random input sample data that meets the distribution characteristics, and the random input sample data is preprocessed.

3. The method for predicting the potential output of renewable energy power generation according to claim 2, characterized in that: For preprocessed data, including: Chi-square test and Kolmogorov-Smirnov test were used to determine whether each factor obeyed the normal distribution or Weibull distribution; If the factors are of known distribution type, the corresponding distribution parameters are estimated by maximum likelihood estimation; According to the probability density function obtained by the estimation, a large number of random samples obeying the distribution are generated by Monte Carlo simulation method for subsequent risk assessment.

4. The method for predicting the potential output of renewable energy power generation according to claim 1, characterized in that: The wavelet analysis of the probability distribution is performed to obtain the output fluctuation characteristics at different time scales, including: The probability density function is discretized and used as the input signal of wavelet transform to perform wavelet packet decomposition to obtain wavelet coefficients at multiple time scales. The abnormal fluctuation threshold of each scale is determined based on historical data. It is judged whether the wavelet coefficient exceeds the threshold to obtain abnormal fluctuation, and the abnormal fluctuation is summarized to construct a feature vector.

5. The method for predicting the potential output of renewable energy power generation according to claim 4, characterized in that: Preprocessing the random input sample data includes: Remove outliers and missing values, and use the maximum and minimum normalization method to normalize the data, map the data to the interval [0, 1], and obtain the preprocessed input sample data; According to the pre-established physical mechanism model of the potential output of renewable energy power generation, the pre-processed input sample data is used as training data, and the mean square error loss function is minimized by the gradient descent algorithm and the back propagation algorithm to optimize the parameters of the physical mechanism model to obtain the trained physical mechanism model; The preprocessed input sample data are respectively substituted into the trained physical mechanism model and the data-driven model, the potential output value corresponding to each input sample is calculated, and the results predicted by the two models are weighted averaged to obtain the final potential output prediction value.

6. The method for predicting the potential output of renewable energy power generation according to claim 1, characterized in that: For photovoltaic power generation, a nonlinear mapping relationship between solar radiation intensity, photovoltaic module conversion efficiency, temperature coefficient, aging attenuation factor and power generation efficiency is established to dynamically correct the predicted value of photovoltaic power generation potential output, including: A deep learning algorithm is used to establish a nonlinear mapping relationship between the factors and the power generation efficiency, and the predicted value is corrected according to the abnormal fluctuation warning results.

7. The method for predicting the potential output of renewable energy power generation according to claim 6, characterized in that: Obtaining a probability density function of the data, and discretizing the probability density function using a histogram method to obtain a discretized probability density function, including: The discretized probability density function is used as an input signal of wavelet transform, and Daubechies wavelet is used to perform multi-scale decomposition to obtain wavelet coefficients at multiple time scales; According to the preset historical data, the abnormal fluctuation threshold of the wavelet coefficient at each time scale is determined; If the wavelet coefficient at the current moment exceeds the abnormal fluctuation threshold, it is determined that abnormal fluctuation exists at the moment; Extract the statistical characteristics of abnormal fluctuations and construct abnormal fluctuation feature vectors; A random forest algorithm is used to classify the abnormal fluctuation feature vector to obtain an abnormality detection model; If the abnormal fluctuation feature vector of the new input data is judged as abnormal by the anomaly detection model, an early warning is triggered.

8. The method for predicting the potential output of renewable energy power generation according to claim 1, characterized in that: For wind power generation, a computational fluid dynamics model is used to simulate the wind field distribution, obtain the wind speed and wind direction parameters at the location of the wind turbine, determine the potential output of wind power in combination with the wind turbine power curve, and make corrections based on the abnormal fluctuation warning results, including: Based on the three-dimensional model of the wind field, the computational fluid dynamics method is used to simulate the speed and direction distribution of the wind field, and the wind speed and wind direction data at the location of the wind turbine are obtained. The wind speed data is input into the power curve of the wind turbine to calculate the theoretical power generation of the wind turbine at the wind speed, and corrections are made according to the abnormal fluctuation warning results.

9. The method for predicting the potential output of renewable energy power generation according to claim 1, characterized in that: Based on the potential output prediction results of photovoltaic and wind power generation, the energy storage capacity configuration and operation strategy are solved through intelligent optimization algorithms to maximize the benefits of the energy storage system, and the dynamic prediction value of the energy storage absorption capacity index is obtained, including: According to the potential output prediction results of photovoltaic and wind power generation, an intelligent optimization algorithm is used to maximize the benefits of the energy storage system while satisfying the constraints of new energy consumption and system stability, and obtain a dynamic prediction value of the energy storage consumption capacity index.

10. The method for predicting the potential output of renewable energy power generation according to claim 1, characterized in that: The adaptive weighted fusion strategy is used to combine the prediction results of the statistical model based on historical data, the dynamic model based on physical mechanism and the data-driven model based on machine learning to predict the energy storage capacity index, including: The weight coefficient of each sub-model is dynamically adjusted according to the prediction error to improve the overall accuracy and robustness of the energy storage capacity index prediction.

Images