A new energy output interval intelligent prediction method based on meteorological changes

Abstract

This invention relates to the field of electronic digital data processing technology and discloses an intelligent prediction method for new energy power output intervals based on meteorological changes. The method includes: acquiring new energy power output time series and meteorological data; determining the number of modes and center frequency; constructing a variable mode decomposition constrained variational model to decompose the power output sequence into multiple modal components; constructing a prediction model for each modal component; dynamically adjusting the network structure and input features of the corresponding model according to real-time meteorological scene labels; superimposing the prediction results of each modal component to obtain the final predicted value; quantifying the error distribution of the predicted value based on the kernel density estimation method; introducing a meteorological sensitivity factor to dynamically adjust the kernel function bandwidth to generate a dynamic probabilistic power output interval; issuing an alarm based on the deviation between the actual power output data monitored by the power grid and the predicted interval, and feeding back the adjusted parameters to optimize the model. This method solves the problems of insufficient accuracy and inadequate utilization of meteorological data in existing technologies, achieving the goal of high accuracy and dynamic interval generation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electronic digital data processing, and particularly relates to a new energy output interval intelligent prediction method based on meteorological changes. BACKGROUND

[0002] Wind power, photovoltaic and other new energy power generation equipment are driven by meteorological factors, and the output has intermittency, volatility and non-stationarity. Power fluctuation will cause problems such as power grid frequency change and insufficient reserve capacity, so it needs to be predicted. The prediction accuracy directly affects the safety and economy of the power grid. High proportion of new energy grid connection needs probability interval prediction, so as to quantify the uncertainty boundary. Deterministic prediction will lead to insufficient robustness of the dispatching model, and the risk of load shedding under extreme scenarios will increase.

[0003] In the existing prediction method, the statistical model (such as the exponential smoothing model): suitable for short-term prediction and wind farms with obvious seasonal regularity of power generation, dependent on meteorological data, difficult to capture the complex nonlinear characteristics of time series model; artificial intelligence methods (such as support vector machine, long short-term memory network): the general implementation process is data cleaning, feature engineering, and prediction output. Among them, data cleaning refers to the detection and repair of outliers, the smoothing processing of noise and the correction of data alignment time sequence. Artificial intelligence methods only rely on feature engineering, and are still disturbed by noise, making the prediction process less adaptable and the optimization and training time-consuming; probability prediction model (such as quantile regression, Bayesian method): probability prediction refers to assuming an error distribution model after the prediction is completed, and obtaining the error range and the probability of the true value falling within the interval. The existing probability error method takes a long time to calculate and has insufficient coverage of risk scenarios.

[0004] For example, the Chinese patent with publication number CN116881627A discloses a microseismic data denoising method based on variational mode decomposition and regularized particle filtering, which provides the following technical solution: step 1, real-time collected microseismic data is subjected to variational mode decomposition; step 2, high-frequency signals are removed based on the peak coefficient, low-frequency IMF components are reconstructed, and a reconstructed signal is obtained; step 3, the reconstructed signal h(t) is filtered based on the regularized particle filtering to obtain a denoising signal. The method disclosed in the present application removes part of the high-frequency components in the microseismic data through the variational mode decomposition technology, retains the low-frequency components close to the signal, and reduces the deviation interference of strong random noise on the particle distribution. The regularized technology is used to correct the depletion problem in particle filtering, improve the filtering accuracy, and enhance the robustness of the algorithm. However, the above-mentioned microseismic data denoising method based on variational mode decomposition and regularized particle filtering cannot realize the probability interval prediction of new energy output, lacks dynamic integration of meteorological factors and high-speed prediction optimization of extreme learning machines, resulting in insufficient prediction accuracy and poor adaptability. SUMMARY

[0005] This invention solves the problems of insufficient prediction accuracy, static intervals, and insufficient utilization of meteorological data in the prior art. It proposes an intelligent prediction method for new energy output intervals based on meteorological changes, achieving the goals of high-precision prediction, dynamic interval generation, and adaptation to extreme weather.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A smart prediction method for new energy output intervals based on meteorological changes, comprising: Obtain the time series of new energy power output and meteorological data, determine the number of modes and center frequency, construct a variable mode decomposition constrained variational model, and decompose the power output sequence into multiple modal components; A prediction model is constructed for each modal component. The network structure and input features of the corresponding model are dynamically adjusted according to the real-time meteorological scene labels. The prediction results of each modal component are superimposed to obtain the final prediction value. Based on the kernel density estimation method, the error distribution of the predicted point values ​​is quantified, and a meteorological sensitive factor is introduced to dynamically adjust the kernel function bandwidth to generate a dynamic probability output interval. An alarm is triggered based on the deviation between the actual power output data monitored by the power grid and the predicted range, and feedback is provided to adjust parameters and optimize the model.

[0007] By integrating meteorological data with power output sequences and employing variable mode decomposition, dynamic prediction models, and kernel density estimation, prediction accuracy can be improved, power output ranges can be dynamically adjusted, and reliable reference data can be provided to the dispatching system. Simultaneously, a closed-loop feedback mechanism enables autonomous model optimization, enhancing the system's adaptability to extreme weather conditions and ensuring power system safety.

[0008] Preferably, the acquisition of new energy output time series and meteorological data specifically includes constructing and normalizing feature vectors of real-time meteorological characteristics and standardizing data scale; the determination of the number of modes and center frequency includes determining the number of decomposed modes based on dynamic meteorological characteristics and modulating the center frequency based on meteorological data so that the center frequency is dynamically correlated with meteorological conditions.

[0009] Normalizing meteorological characteristics and standardizing data scales helps eliminate the influence of dimensions and improves model training stability. Dynamically modulating the center frequency based on meteorological data makes the decomposition process more closely match actual meteorological conditions, improving the accuracy and adaptability of mode decomposition.

[0010] Preferably, the construction of the variable mode decomposition constrained variational model specifically includes establishing a constrained variational model of variable mode decomposition, with the objective of minimizing the sum of the bandwidths of all mode functions, the constraint condition being that the sum of all mode functions equals the original output sequence, and adding a regularization term of meteorological characteristics for constraint.

[0011] Minimizing modal bandwidth makes each mode more compact in the frequency domain, reducing mode aliasing. Adding meteorological feature regularization constraints makes the decomposition process take into account the influence of meteorological factors, improving the decomposition's relevance and accuracy.

[0012] Preferably, the construction of a prediction model for each modal component specifically includes constructing an extreme learning machine model with an input layer, a hidden layer, and an output layer for each modal component. First, the parameters of the hidden nodes are randomly set, then the output matrix of the hidden layer is calculated, and finally the output weight matrix is ​​calculated.

[0013] Extreme Learning Machine (ELM) models possess rapid learning capabilities, efficiently capturing dynamic features of subsequences. By directly calculating the output weight matrix, it avoids iterative optimization, significantly accelerating training speed and reducing computational resource consumption.

[0014] Preferably, the method of quantifying the error distribution of point prediction values ​​based on kernel density estimation specifically includes: establishing a random distribution model of the new energy output range using kernel density estimation, selecting a Gaussian kernel as the kernel function; and calculating the bandwidth parameter of kernel density estimation based on the point prediction error data.

[0015] Kernel density estimation methods can flexibly model error distributions without assuming a fixed distribution form. Choosing a Gaussian kernel simplifies the calculation process, facilitates the generation of probabilistic output intervals, and improves the reliability of interval predictions.

[0016] Preferably, the alarm based on the deviation between the actual power output data monitored by the power grid and the prediction range specifically includes: monitoring the actual power output data of the power grid in real time, calculating the deviation between it and the prediction range, and triggering an alarm mechanism to perform anomaly checks when the deviation exceeds a preset threshold.

[0017] Real-time monitoring and deviation alarm mechanisms can promptly identify predicted anomalies and trigger inspection processes. This helps to quickly respond to output fluctuations, optimize model parameters, and improve system robustness.

[0018] Preferably, the process of decomposing the output sequence into multiple modal components includes iteratively solving a constrained variational problem, using a quadratic penalty term and the Lagrange multiplier method, introducing a Lagrange augmented function, and using the alternating direction multiplier method to solve the optimization problem; alternately updating the modal functions and center frequencies until the accuracy requirements are met, and finally obtaining the decomposed multiple modal components.

[0019] Iterative solution methods ensure that the decomposition process converges to the optimal solution, improving the accuracy of modal components. The alternating direction multiplier method efficiently handles constrained problems, reducing computational complexity and guaranteeing decomposition efficiency.

[0020] Preferably, the method of introducing meteorological sensitive factors to dynamically adjust the kernel function bandwidth specifically includes introducing meteorological sensitive factors as bandwidth adjustment coefficients, dynamically taking values ​​according to real-time meteorological scenarios, taking a baseline value under normal weather conditions, and increasing the factor value to expand the bandwidth under extreme weather scenarios; when generating dynamic probability output intervals, calculating the percentiles of the error distribution at each point based on the updated bandwidth parameters to form a prediction interval with adjustable confidence levels.

[0021] The bandwidth of meteorological sensitive factors is dynamically adjusted to adapt the error distribution to different weather conditions, expanding the range under extreme weather conditions to cover higher risks. This generates a dynamically adjustable forecast range, improving the accuracy of uncertainty quantification.

[0022] Preferably, the feedback adjustment parameter optimization model includes, based on alarm information, adjusting the parameters in the variable mode decomposition model including the number of modes, center frequency, and constraint regularization term, adjusting the network structure or input features in the prediction model, adjusting the meteorological sensitivity factor or bandwidth parameter in the kernel density estimation, and optimizing the error distribution.

[0023] The feedback adjustment mechanism enables adaptive optimization of model parameters, continuously improving prediction performance based on actual results. This enhances the model's generalization ability and long-term stability under variable weather conditions.

[0024] Preferably, the step of dynamically adjusting the network structure and input features of the corresponding model according to the real-time weather scene labels includes: in typhoon weather scenarios, increasing the hidden layer to three layers to enhance the ability to capture complex patterns and features, and adding input features including short-term wind speed fluctuation rate and wind direction change rate; in rainy weather scenarios, adding humidity features to the input nodes; in sunny weather scenarios, reducing the number of hidden nodes; and finally, superimposing the predicted values ​​of each modal component and performing inverse normalization to obtain the final predicted value.

[0025] Dynamically adjusting the network and input features allows the model to optimize its structure for different weather scenarios. For example, it increases complexity to capture drastic changes during typhoons, while simplifying the structure to improve efficiency during sunny weather. This improves prediction accuracy and speed while reducing resource waste.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows.

[0027] 1. This invention integrates meteorological data with new energy output sequences, employs variable mode decomposition (VMD) technology, determines the number of modes and center frequencies based on dynamic meteorological characteristics, and adds meteorological regularization constraints to make the decomposition process more closely reflect actual meteorological conditions. This effectively reduces aliasing between different modes, making each mode more compact in the frequency domain, thereby improving the accuracy and adaptability of time-series data decomposition. Through iterative solution of the optimization problem, it ensures that the decomposition results converge to the optimum, providing cleaner and more characteristic input data for subsequent prediction models, laying the foundation for high-precision prediction.

[0028] 2. This invention employs dynamic prediction models such as Extreme Learning Machines (ELM) to rapidly learn each modal component and dynamically adjust the network structure and input features based on real-time meteorological scene labels. Hidden layers and input features are added during typhoon weather to capture complex patterns, while the structure is simplified during clear weather to improve efficiency. This adaptive adjustment mechanism not only improves the model's prediction accuracy for non-stationary power output sequences but also significantly accelerates the training process, avoids iterative optimization, and reduces computational resource consumption, thereby achieving efficient and reliable power output prediction.

[0029] 3. This invention quantifies the distribution of prediction errors based on the kernel density estimation method, introduces meteorological sensitivity factors to dynamically adjust the kernel function bandwidth, and generates a dynamic probability output range. Combined with real-time power grid monitoring data, an alarm mechanism is triggered when the deviation exceeds a threshold, and model parameters are adjusted through closed-loop feedback. This enhances the system's adaptability to extreme weather conditions, improves the accuracy of uncertainty quantification, ensures the safety and stability of the power system, and provides a reliable reference for dispatching decisions. Attached Figure Description

[0030] Figure 1 This is an overall flowchart of the intelligent prediction method for new energy output range based on meteorological changes according to the present invention.

[0031] Figure 2 This is a flowchart illustrating the intelligent prediction method for new energy output ranges based on meteorological changes, as described in this invention.

[0032] Figure 3 This is a hierarchical structure diagram of the extreme learning machine for an intelligent prediction method of new energy output range based on meteorological changes, according to the present invention. Detailed Implementation

[0033] See Figures 1-3 As shown, a smart prediction method for new energy output range based on meteorological changes includes: Obtain the time series of new energy power output and meteorological data, determine the number of modes and center frequency, construct a variable mode decomposition constrained variational model, and decompose the power output sequence into multiple modal components; A prediction model is constructed for each modal component. The network structure and input features of the corresponding model are dynamically adjusted according to the real-time meteorological scene labels. The prediction results of each modal component are superimposed to obtain the final prediction value. Based on the kernel density estimation method, the error distribution of the predicted point values ​​is quantified, and a meteorological sensitive factor is introduced to dynamically adjust the kernel function bandwidth to generate a dynamic probability output interval. An alarm is triggered based on the deviation between the actual power output data monitored by the power grid and the predicted range, and feedback is provided to adjust parameters and optimize the model.

[0034] like Figures 1-3 In one embodiment shown, Figure 1 This is an overall flowchart of the intelligent prediction method for new energy output range based on meteorological changes according to the present invention. Figure 2 This is a flowchart illustrating the intelligent prediction method for new energy output ranges based on meteorological changes, as described in this invention. Figure 3 This is a hierarchical structure diagram of the extreme learning machine for an intelligent prediction method of new energy output intervals based on meteorological changes, as proposed in this invention. First, the time series of new energy output and meteorological data are acquired, and feature vectors are constructed and normalized for real-time meteorological characteristics to standardize the data scale. The number of modes is determined based on dynamic meteorological characteristics, and the center frequency is modulated based on the meteorological data to dynamically correlate the center frequency with meteorological conditions. Then, a variable mode decomposition constrained variational model is constructed, with the objective of minimizing the sum of the bandwidths of all mode functions. The constraint condition is that the sum of all mode functions equals the original output sequence, and a regularization term for meteorological characteristics is added for constraint. The constrained variational problem is solved iteratively, using a quadratic penalty term and the Lagrange multiplier method, introducing the Lagrange augmented function, and employing the alternating direction multiplier method to solve the optimization problem. The mode functions and center frequency are alternately updated until the accuracy requirements are met, ultimately decomposing the output sequence into multiple modal components.

[0035] Next, a prediction model is constructed for each modal component, specifically using an extreme learning machine model with input, hidden, and output layers. First, the parameters of the hidden nodes are randomly set, then the output matrix of the hidden layer is calculated, and finally the output weight matrix is ​​calculated. The network structure and input features of the corresponding model are dynamically adjusted according to the real-time weather scene labels. For example, in the typhoon weather scene, the hidden layer is increased to three layers to enhance the ability to capture complex patterns and features, and input features including short-term wind speed fluctuation rate and wind direction change rate are added. In the rainy weather scene, humidity features are added, and in the sunny weather scene, the number of hidden nodes is reduced. The prediction results of each modal component are superimposed and inversely normalized to obtain the final prediction value.

[0036] Then, based on the kernel density estimation method, the error distribution of the point prediction values ​​is quantified, and a random distribution model of the new energy output interval is established using the kernel density estimation method. A Gaussian kernel is selected as the kernel function, and the bandwidth parameter of the kernel density estimation is calculated based on the point prediction error data. A meteorological sensitivity factor is introduced to dynamically adjust the kernel function bandwidth. The meteorological sensitivity factor is used as the bandwidth adjustment coefficient and is dynamically selected according to the real-time meteorological scenario. The baseline value is taken under normal weather conditions, and the factor value is increased to expand the bandwidth under extreme weather scenarios, thereby generating a dynamic probability output interval. Based on the updated bandwidth parameter, the percentile of the error distribution of each point is calculated to form a prediction interval with adjustable confidence level.

[0037] Finally, an alarm is triggered based on the deviation between the actual power output data monitored by the power grid and the prediction interval. The actual power output data of the power grid is monitored in real time, and the deviation between it and the prediction interval is calculated. When the deviation exceeds a preset threshold, an alarm mechanism is triggered to check for anomalies. Feedback is also provided to adjust the parameters and optimize the model. Based on the alarm information, the number of modes, center frequency and constraint regularization parameters in the variable mode decomposition model are adjusted, the network structure or input features in the prediction model are adjusted, and the meteorological sensitivity factor or bandwidth parameter in the kernel density estimation is adjusted to optimize the error distribution.

[0038] In another embodiment, the present invention provides a method for intelligent prediction of new energy power output and power output range, comprising the following steps: S110: Obtain the time series of new energy output, determine the number of modes and center frequency based on meteorological factor input, establish a constrained variational model for variable mode decomposition, and reduce the aliasing rate of different modes.

[0039] First, feature vectors are constructed and normalized for real-time meteorological features.

[0040] The number of decomposition modes is determined based on dynamic meteorological characteristics. Unlike the center frequency of random modes in general variable mode decomposition, the center frequency is modulated based on meteorological data.

[0041] The core optimization problem of variable mode decomposition is established. The objective is to minimize the sum of the bandwidths of all mode functions, i.e., to make each mode more compact in the frequency domain. The constraint is that the sum of all mode functions equals the original signal f(x). Regularization terms based on meteorological characteristics are added to the constraints.

[0042] Then, iterative solutions are performed. The constrained variational problem is transformed into an unconstrained variational problem, and a Lagrange augmented function is introduced using a quadratic penalty term and the Lagrange multiplier method. By finding the saddle point of the augmented Lagrange L, the optimization problem is solved using the alternating direction multiplier method. Therefore, solving the problem only requires updating u. k and w k The value of u is obtained by solving for u. k The optimization problem is used to update the parameters each time.

[0043] For each mode, during the initialization of u k and w k After obtaining the value, update u according to the above formula. k and w k The value of λ is then updated according to the algorithm until the accuracy is satisfied, thus obtaining different modes and their center frequencies.

[0044] Specifically, the time series of new energy output is obtained, the number of modal decomposition levels is specified, a constrained variational model of variable modal decomposition is established, and the model is solved. The constrained variational model can be expressed in the following form: The goal of a constrained variational model is to minimize the set of modes u. k and center frequency set w k The expression is a summation term for all modality indices k, each summation term involving modality u. k (t) is the squared L2 norm after a series of operations. The specific operations include: first calculating the partial derivative, then convolving it with a function consisting of the Dirac function δ(t) and the imaginary unit j divided by π multiplied by t, and finally multiplying it by the negative j of the exponential term e multiplied by w. k Multiply by t. The constraint is all modes u. k The sum of (t) equals the original signal f(t).

[0045] In this case, to solve the constrained variational problem, we transform it into an unconstrained variational problem by using a quadratic penalty term and the Lagrange multiplier method, introducing a Lagrange augmented function. The transformed unconstrained optimization problem takes the form: The transformed unconstrained optimization problem takes the form of a Lagrange augmented function L, which consists of three parts: the first part is the parameter a multiplied by a summation term similar to that in a constrained variational model, which involves each mode u. k The L2 norm squared after the above operation on (t); the second part is the original signal f(t) minus all modes u k The sum of L2 norm squares of (t); the third part is the Lagrange multiplier λ(t) and the constraints (i.e., f(t) minus all u) k The inner product of (t) and (t). Where a is the penalty parameter and λ is the Lagrange multiplier.

[0046] The optimization problem is solved by finding the saddle point of the augmented Lagrange L and using the alternating direction multiplier method. Therefore, solving the problem only requires updating u. k and w k The value of u is obtained by solving for u. k The optimization problem is used to update the parameters each time.

[0047] For all w, use u for the nth time.k Solve for u in the (n+1)th iteration k The solution formula is: At its core, it is a minimization problem, that is, finding the u that minimizes the objective function. k The objective function consists of the sum of two terms. The first term is the parameter a multiplied by the square of the L2 norm of the aforementioned Dirac function δ(t). The second term is the square of another L2 norm, which is the original signal f(t) minus all modes u. k The expression obtained by adding the sum of (t) and the Lagrange multiplier λ(t) divided by 2.

[0048] The modal functionals of this non-optimal constraint problem are updated as follows: In the frequency domain, the mode u of the (n+1)th iteration k The Fourier transform value is equal to the Fourier transform value of the original signal f minus the Fourier transform value of all other modes u. i The sum of the Fourier transform values ​​(where u is not equal to k), plus the Fourier transform value of the Lagrange multiplier λ divided by 2, the whole result divided by 1 plus 2a multiplied by (frequency w minus center frequency w) k The square of ).

[0049] Based on this, the optimization problem with respect to is solved using a formula similar to the one described above, written in the frequency domain as follows: w in the (n+1)th iteration k The value is equal to w multiplied by the mode u over the range from frequency 0 to infinity. k Integrate the modulus square of the Fourier transform, then divide by the modulus u over the frequency range from 0 to infinity. k Integrate the modulus square of the Fourier transform.

[0050] S120: An extreme learning machine model is built for each modal component to make predictions, utilizing its fast learning characteristics to capture dynamic features of subsequences. Furthermore, the prediction network is adjusted in real time based on the meteorological scene labels. For example... Figure 2 As shown, Figure 2 This is a flowchart illustrating the intelligent prediction method for new energy output range based on meteorological changes, as described in this invention. The core idea is to learn each mode, and finally perform standardized synthesis to determine the final output range.

[0051] For each modal component, construct as follows Figure 3 The extreme learning machine model shown makes predictions, utilizing its fast learning characteristics to capture dynamic features of subsequences. Figure 3 This is a hierarchical structure diagram of the Extreme Learning Machine (ELM) for an intelligent prediction method of new energy output range based on meteorological changes, as proposed in this invention. The ELM has a three-layer structure: an input layer, a hidden layer, and an output layer.

[0052] The training process can be summarized as follows: First, the hidden node parameters are randomly set; second, the hidden layer output matrix H is calculated; finally, the output weight matrix β is calculated to be equal to H*O, where H* is the Moore-Penrose generalized inverse matrix of H. This avoids iterative optimization and greatly accelerates training.

[0053] For different weather scene labels, the network structure is adjusted and different activation functions are selected. In particular, for scenes with rapidly changing winds, such as typhoons, the hidden layers are increased to 3 layers to enhance the model's ability to capture complex patterns and features; humidity features are added to the input nodes for rainy days; and the number of hidden nodes is reduced for sunny days to improve the model's prediction speed and reduce computational resource consumption.

[0054] Finally, the predicted values ​​of each modality are superimposed and inversely normalized to obtain the final predicted value.

[0055] S130: Quantify the prediction error distribution based on kernel density estimation and generate confidence intervals. The error distribution bandwidth is dynamically adjusted to account for meteorological factors.

[0056] A stochastic distribution model for the renewable energy output range is established using the kernel density estimation method. Then, the potential probability density function for the renewable energy output range is derived.

[0057] Choosing a Gaussian kernel as the kernel function allows us to represent the potential probability density function, thereby updating the form of the potential probability density function for the new energy output range. Based on the error distribution parameters at each point, the set of these parameters yields the prediction range.

[0058] The bandwidth factor is increased by adding a meteorological sensitivity factor to the generated bandwidth. The value of the meteorological sensitivity factor varies under different weather scenarios, and it is increased in extreme weather to cover extreme power output fluctuations.

[0059] S140: Based on the power output data and fluctuations monitored by the power grid, issue alarms to the meteorological monitoring system, thereby adjusting the parameters of each link and optimizing the prediction model.

[0060] This invention primarily considers model deployment in wind farms. Under both normal and extreme weather conditions, the input to the training prediction model includes two types of data: historical power output data and meteorological feature data. The selection, weighting, and model structure of the meteorological features are dynamically adjusted based on the weather scenario labels.

[0061] In one embodiment, under normal weather conditions without extreme weather events, the system employs a general forecast. The input data for this general forecast is based on historical power output time-series data from the wind farm over the past year, as well as conventional meteorological elements provided by weather stations or numerical weather prediction, including wind speed, wind direction, temperature, humidity, and air pressure. These basic meteorological parameters, together with the historical power output data, constitute the core feature set. Specific parameters include: There are two hidden layers; First hidden layer nodes 20-40; The activation function is ReLU; In the feature weighting, wind speed and historical power output have relatively large weights; The probability interval meteorological sensitivity factor is 1.0, and the interval width is 20%-30% of the rated power.

[0062] In another embodiment, when the system detects wind speeds consistently exceeding 17.2 m / s, a significant increase in pressure gradient, and a typhoon warning issued by the meteorological department, it automatically switches to "typhoon weather." At this time, the system will activate forecasts specifically optimized for typhoon characteristics. In addition to the historical power output data and basic meteorological elements relied upon, significantly more input parameters closely related to typhoon dynamics are added. These new parameters include: short-term wind speed fluctuation rate, rapid change rate of wind direction, typhoon center pressure value, surrounding pressure gradient, relative position and distance between the wind farm and the typhoon center, and real-time rainfall intensity. By introducing these highly dynamic characteristics, the system can more precisely characterize the extreme wind field structure, intense turbulence effects, and dramatic weather changes within and around the typhoon vortex, thereby making more reliable forecasts under extreme wind conditions and providing crucial support for safety assurance and dispatch decisions. Specific parameters include: The number of hidden layers is three; Add nodes (nodes 50-80 in the first hidden layer); The activation function is PReLU; The weights for wind direction change rate, pressure gradient, and relative position in the feature weights have been increased. The probability interval meteorological sensitivity factor is 3.5, and the interval width is 60%-80% of the rated power.

[0063] When clear weather occurs with extremely high temperatures (above 40°C) and extremely low wind speeds (below 2 m / s), the system classifies it as an "extremely hot and windless weather" scenario. Under these conditions, although solar radiation is strong, wind power generation capacity is almost completely lost. For this specific scenario, the input parameters, in addition to basic meteorological data, will specifically enhance temperature and its derived thermodynamic parameters, and incorporate solar radiation intensity data. Specific parameters include: The number of hidden layers is three; Reduce the number of nodes (30-50 nodes in the first hidden layer); The activation function is SELU; The weighting of features includes perceived temperature, cloud cover, solar radiation intensity, and grid load. Meteorological sensitivity factor 0.8, range width (5%-10% of rated power).

[0064] This invention significantly improves the adaptability and operational efficiency of the power system in the face of weather changes by integrating meteorological data with the entire process of new energy output forecasting. Traditional methods often struggle to capture the inherent frequency domain characteristics of non-stationary new energy output sequences, leading to insufficient prediction accuracy. Furthermore, probability interval predictions are often based on fixed error distribution assumptions, failing to dynamically respond to the uncertainties brought about by extreme weather. This invention, however, first introduces a variable mode decomposition technique modulated by meteorological factors in the data preprocessing stage. By dynamically determining the number of modes and the center frequency, it effectively reduces the aliasing phenomenon between different modes, enabling clear separation and enhancement of the frequency domain characteristics of the output sequence. This not only overcomes the poor adaptability of traditional decomposition methods to non-stationary sequences but also provides richer and higher-quality feature inputs for subsequent prediction models.

[0065] Subsequently, an Extreme Learning Machine (ELM) prediction model was constructed for each modal component. Leveraging its rapid learning characteristics, the training time was significantly shortened, avoiding the computational resource consumption caused by complex iterative optimization. Simultaneously, the network structure parameters and input features were dynamically adjusted based on real-time weather scene labels. For example, the network depth was increased to capture complex patterns during typhoon weather, while the structure was simplified to improve efficiency during clear weather. This dynamic adaptation mechanism ensured the model's stability and accuracy under different weather conditions. In the error processing stage, a kernel density estimation method was used to quantify the prediction error distribution, and a meteorological sensitivity factor was introduced to dynamically adjust the kernel function bandwidth. This ensured that the generated probability output interval could reflect the impact of weather changes on uncertainty in real time, especially automatically expanding the interval range under extreme weather conditions. This provided a more comprehensive coverage of potential risks and offered a more reliable decision-making basis for power grid dispatching.

[0066] Furthermore, this invention also constructs a closed-loop optimization mechanism. By monitoring the deviation between the actual power output data of the power grid and the prediction range in real time, it triggers alarms and feeds back to adjust model parameters, such as the constraint terms of variable mode decomposition, the structure of the prediction network, or the bandwidth parameters of kernel density estimation. This enables the model to learn autonomously and continuously improve. This self-optimization capability not only reduces long-term maintenance costs but also further consolidates the stability and foresight of the prediction system.

[0067] Overall, this invention deeply integrates meteorological data into all aspects of the forecasting process, achieving dynamic optimization across the entire chain from data decomposition and model prediction to error assessment. This not only significantly improves the accuracy and speed of forecasting new energy output points and interval forecasts, but also enhances the power system's resilience to meteorological changes, effectively reduces grid operation risks, and ensures the security and economy of energy supply, providing strong technical support for the stable operation of the power system under high-proportion new energy access.

Claims

1. A method for intelligent prediction of new energy output ranges based on meteorological changes, characterized in that, include: Obtain the time series of new energy power output and meteorological data, determine the number of modes and center frequency, construct a variable mode decomposition constrained variational model, and decompose the power output sequence into multiple modal components; A prediction model is constructed for each modal component. The network structure and input features of the corresponding model are dynamically adjusted according to the real-time meteorological scene labels. The prediction results of each modal component are superimposed to obtain the final prediction value. Based on the kernel density estimation method, the error distribution of the predicted point values ​​is quantified, and a meteorological sensitive factor is introduced to dynamically adjust the kernel function bandwidth to generate a dynamic probability output interval. An alarm is triggered based on the deviation between the actual power output data monitored by the power grid and the predicted range, and feedback is provided to adjust parameters and optimize the model.

2. The intelligent prediction method for new energy output range based on meteorological changes according to claim 1, characterized in that, The acquisition of new energy output time series and meteorological data specifically includes constructing and normalizing feature vectors of real-time meteorological characteristics and standardizing data scale; the determination of the number of modes and center frequency includes determining the number of decomposed modes based on dynamic meteorological characteristics and modulating the center frequency based on meteorological data to make the center frequency dynamically correlated with meteorological conditions.

3. The intelligent prediction method for new energy output range based on meteorological changes according to claim 2, characterized in that, The construction of the variable mode decomposition constrained variational model specifically includes establishing a constrained variational model of variable mode decomposition, with the objective of minimizing the sum of the bandwidths of all mode functions, the constraint condition being that the sum of all mode functions equals the original output sequence, and adding a regularization term of meteorological characteristics for constraint.

4. The intelligent prediction method for new energy output range based on meteorological changes according to claim 3, characterized in that, The construction of a prediction model for each modal component specifically includes constructing an extreme learning machine model with an input layer, a hidden layer, and an output layer for each modal component. First, the parameters of the hidden nodes are randomly set, then the output matrix of the hidden layer is calculated, and finally the output weight matrix is ​​calculated.

5. The intelligent prediction method for new energy output range based on meteorological changes according to claim 4, characterized in that, The method of quantifying the error distribution of point prediction values ​​based on kernel density estimation specifically includes: establishing a random distribution model of the new energy output range using kernel density estimation, selecting a Gaussian kernel as the kernel function; and calculating the bandwidth parameter of kernel density estimation based on the point prediction error data.

6. The intelligent prediction method for new energy output range based on meteorological changes according to claim 5, characterized in that, The alarm mechanism based on the deviation between the actual power output data monitored by the power grid and the predicted range specifically includes: real-time monitoring of the actual power output data of the power grid, calculating the deviation between it and the predicted range, and triggering an alarm mechanism to perform anomaly checks when the deviation exceeds a preset threshold.

7. A method for intelligent prediction of new energy output range based on meteorological changes according to claim 2 or 6, characterized in that, The process of decomposing the power output sequence into multiple modal components involves iteratively solving a constrained variational problem, using a quadratic penalty term and the Lagrange multiplier method, introducing a Lagrange augmented function, and using the alternating direction multiplier method to solve the optimization problem; alternately updating the modal functions and center frequencies until the accuracy requirements are met, and finally obtaining the decomposed multiple modal components.

8. A method for intelligent prediction of new energy output range based on meteorological changes according to claim 5 or 6, characterized in that, The specific steps of introducing meteorological sensitive factors to dynamically adjust the kernel function bandwidth include: introducing meteorological sensitive factors as bandwidth adjustment coefficients, dynamically taking values ​​based on real-time meteorological scenarios, taking a baseline value under normal weather conditions, and increasing the factor value to expand the bandwidth under extreme weather scenarios; when generating dynamic probability output intervals, calculating the percentiles of the error distribution at each point based on the updated bandwidth parameters to form a prediction interval with adjustable confidence levels.

9. The intelligent prediction method for new energy output range based on meteorological changes according to claim 6, characterized in that, The feedback adjustment parameter optimization model includes adjusting the parameters in the variable mode decomposition model, including the number of modes, center frequency, and constraint regularization term, based on alarm information; adjusting the network structure or input features in the prediction model; adjusting the meteorological sensitivity factor or bandwidth parameter in the kernel density estimation; and optimizing the error distribution.

10. A method for intelligent prediction of new energy output range based on meteorological changes according to claim 4 or 6, characterized in that, The method of dynamically adjusting the network structure and input features of the corresponding model based on real-time weather scene labels includes: in typhoon weather scenarios, increasing the hidden layer to three layers to enhance the ability to capture complex patterns and features, and adding input features including short-term wind speed fluctuation rate and wind direction change rate; in rainy weather scenarios, adding humidity features to the input nodes; in sunny weather scenarios, reducing the number of hidden nodes; and finally, superimposing the predicted values ​​of each modal component and performing inverse normalization to obtain the final predicted value.

Images