A new energy power generation prediction method and system based on numerical weather prediction

Abstract

The application provides a new energy power generation prediction method and system based on numerical weather prediction, and relates to the technical field of new energy power generation. First, the numerical weather prediction spatial grid multi-element data of a target area is obtained, feature construction is performed, and time domain features and space domain features are obtained. The time domain features and the space domain features are spliced to obtain space-time features. A preset point prediction model and a quantile regression model are trained to obtain trained point prediction models and quantile regression models. Finally, the space-time features are input into the trained point prediction models and quantile regression models, and new energy power generation prediction results are output. The application improves the accuracy of new energy power generation prediction by extracting the space-time features of the numerical weather prediction spatial grid multi-element data.

Description

Technical Field

[0001] This invention belongs to the field of new energy power generation, and more specifically relates to a new energy power generation prediction method and system based on numerical weather forecasting. Background Technology

[0002] Against the backdrop of the global energy transition and the ongoing pursuit of "dual carbon" goals, the penetration rate of new energy sources such as wind power and photovoltaics in the power system continues to increase. However, the output of new energy sources is significantly affected by meteorological conditions, exhibiting randomness and intermittency, which poses serious challenges to the safe and stable operation of the power system, dispatch planning, reserve capacity allocation, and electricity market transactions. Therefore, achieving short-term, high-precision point and probabilistic forecasting of new energy power generation has become a key technological support for improving the flexibility of the power system and promoting the efficient consumption of new energy.

[0003] Current mainstream methods for predicting renewable energy power generation still have significant limitations. First, most methods rely solely on single-point meteorological forecast data for the target power plant or construct spatiotemporal autoregressive models based solely on on-site measurement data, failing to fully leverage the gridded information provided by numerical weather prediction over a larger spatial area. This results in insufficient utilization of the spatial correlations of meteorological elements. For key influencing factors such as cloud cover, wind field, and multi-level meteorological parameters, existing methods struggle to effectively capture their spatial variability and temporal consistency. Under complex meteorological scenarios such as intermittent cloudy conditions, frontal passages, and wind shear, prediction accuracy and the ability to characterize probability distributions significantly decrease.

[0004] Existing methods have improved prediction performance by introducing more advanced statistical learning frameworks, such as numerical weather prediction data update methods, source-grid-load-storage system coordinated control technology, and renewable energy power plant output prediction devices. However, a generalized automatic extraction and analysis framework for numerical weather prediction grid information is still lacking. Existing methods often cannot simultaneously meet the dual requirements of point prediction and probabilistic prediction, making it difficult to achieve compatibility and transfer in multi-energy scenarios such as wind power and photovoltaics. Especially when the training sample is limited, the robustness and generalization ability of the model are still insufficient, resulting in poor accuracy of renewable energy power generation prediction results, which restricts its widespread application in actual power system operation. Summary of the Invention

[0005] To address the problem of poor accuracy in existing new energy power generation forecasting, this invention provides a new energy power generation forecasting method and system based on numerical weather forecasting, which improves the accuracy of new energy power generation forecasting.

[0006] To achieve the above-mentioned technical effects, the technical solution of the present invention is as follows: S1: Acquire multi-element data of the numerical weather prediction spatial grid for the target area; S2: Construct features from multi-factor data to obtain time-domain features and spatial-domain features; S3: Combine the time-domain features and spatial-domain features to obtain spatiotemporal features; S4: Train the preset point prediction model and the preset quantile regression model to obtain the trained point prediction model and quantile regression model. S5: Input the spatiotemporal features into the trained point prediction model and quantile regression model, and output the prediction results of new energy power generation.

[0007] Furthermore, the multi-element data of the numerical weather forecast spatial grid for the target area includes: photovoltaic power generation data and wind power generation data; Photovoltaic power generation data is surface shortwave irradiance. Temperature at a height of 2 meters Low cloud cover Medium cloud cover High cloud cover Total cloud cover ; Wind power generation data consists of the east-west wind speed components of the multi-mode layer. East-west wind speed component Wind speed modulus and wind direction .

[0008] Furthermore, step S1 also includes: Acquire historical power data for the target area, and align the timestamps of the historical power data for the target area with the multi-element data of the numerical weather prediction spatial grid; fill in missing values ​​in the historical power data, and identify and remove outliers in the historical power data. Based on historical surface shortwave irradiance data and historical power data, a clear sky baseline curve is generated. Based on the clear sky baseline curve, the historical power data and surface shortwave irradiance data of the target area are normalized to obtain the normalized power index characterizing weather fluctuations.

[0009] Furthermore, the process of obtaining the time-domain features is as follows: Key meteorological variables were selected from multi-element data of numerical weather prediction spatial grids based on the type of new energy source. x ; Set up a sliding window to extract key meteorological variables. x Calculate key meteorological variables x The variance of the time series within the sliding window is expressed as:

[0010] In the formula, express The time series variance at time t. t Let k represent the current time, and k represent the prediction lead time. Indicates the target prediction time. Indicates the length of the sliding window. i This represents the time index of the time window. This represents the arithmetic mean of the values ​​of numerical weather prediction variables. This represents the value of the numerical weather forecast variable at time i. Based on key meteorological variables x， Construct time-lag features that include historical values ​​several moments before the prediction time and advance features that include predicted values ​​several moments after the prediction time; For the same forecast time, several key meteorological variables are obtained, and the multi-time fusion feature is calculated. The expression is as follows:

[0011] In the formula, Indicates multi-temporal fusion features, Indicates weight, Indicates the dth time Forecast value for the time; Based on the time series variance, time lag features, lead features, and multi-time fusion features, time domain features are obtained.

[0012] Furthermore, the process of obtaining spatial domain features is as follows: Based on multi-factor data, at the prediction time Calculate the spatial standard deviation for each spatial grid, expressed as:

[0013] In the formula, Indicates spatial standard deviation. This represents the numerical weather forecast variable values ​​for i spatial grids. Indicates the number of spatial grids. express The mean value of all spatial grid numerical weather forecast variables at any given time; Centered on the target site in the target area, the spatial grid is divided into L layers. The smooth mean of all grids in each layer is calculated. The smooth mean of all layers is weighted and fused to obtain the smooth feature. The expression for calculating the smoothed mean is as follows:

[0014] In the formula, express The smoothed mean at time t, Represents the set of grids in the i-th layer; express At time j, the numerical weather forecast variable value of the j-th grid; The calculation expression for the smoothing feature is as follows:

[0015] In the formula, express Smoothness characteristics of time, Indicates the weight of the i-th layer; Based on multi-element data from the spatial grid of numerical weather prediction, historical event sequence data of all grids are extracted to construct a historical spatiotemporal data matrix. ;in, For the number of time samples, The number of spatial grid points is given; the historical spatiotemporal data matrix is ​​standardized to obtain a standard historical spatiotemporal data matrix; singular value decomposition is performed on the standard historical spatiotemporal data matrix; the first r principal components are selected; the original high-dimensional spatial data is projected onto the low-dimensional principal component space to obtain the dimensionality-reduced principal component features. Based on the spatial standard deviation, smoothing features, and principal component features, spatial domain features are obtained.

[0016] Furthermore, the process of step S3 is as follows: By concatenating the time-domain features and spatial-domain features, a set of candidate features is obtained. Calculate the Pearson correlation coefficients of all features in the candidate feature set and the acquired historical power data, and delete features whose Pearson correlation coefficients are lower than a preset threshold to obtain the first feature subset. Calculate the variance inflation factor among all features in the first feature subset, delete features whose variance inflation factor value is higher than a preset factor threshold, and obtain the second feature subset; Based on the preset gradient boosting tree base learner, the importance scores of the features in the second feature subset are calculated and sorted from high to low according to the importance scores. Features with importance scores lower than the preset importance threshold are deleted to obtain the spatiotemporal features.

[0017] Further, step S4 specifically includes: Based on time-domain and spatial-domain features, a candidate model library containing N models is constructed. The database includes: a time feature model with input features of time lag and lead features, a time enhancement model with input features of time-series fusion features and time series variance, a spatial feature model with input features of spatial standard deviation and smoothing features, a spatial dimensionality reduction model with input features of principal component features, and a fusion model with input features of spatial domain and time domain features. Each model in the training candidate model library has a preset point prediction model and a preset quantile regression model; Calculate the comprehensive index of each model in the candidate model library, and select the model with the smallest comprehensive index as the optimal model; The preset point prediction model and preset quantile regression model of the optimal model are used as the trained point prediction model and quantile regression model.

[0018] Furthermore, the process of training the preset point prediction model and preset quantile regression model for each model in the candidate model library is as follows: During the training of the point prediction model, a point prediction loss function is constructed. When the loss function converges, the trained point prediction model is obtained. The point prediction loss function is either a squared loss function or an absolute bias loss function. The expression for the squared loss function is:

[0019] In the formula, Let M represent the squared loss function, M represent the total number of training samples, and m represent the index of the training sample. This represents the true value of the m-th sample. This represents the model prediction value for the m-th sample;

[0020] In the formula, Let M represent the absolute bias loss function, M represent the total number of training samples, and m represent the index of the training sample. This represents the true value of the m-th sample. This represents the model prediction value for the m-th sample; During the training of the quantile regression model, a quantile loss function is constructed. When the loss function converges, the trained quantile regression model is obtained. The expression for the quantile loss function is:

[0021] In the formula, Let M represent the quantile loss function, M represent the total number of training samples, and m represent the index of the training sample. This represents the true value of the m-th sample. Indicates fractional place The predicted value, This represents the quantile loss kernel function.

[0022] Furthermore, the calculation expression for the comprehensive index is as follows:

[0023] In the formula, express, Indicates the mean absolute error. This represents the root mean square error. This represents the root mean square error. Indicates the weighting coefficient. Indicates the weighting coefficient. This represents the weighting coefficient.

[0024] This invention also provides a new energy power generation forecasting system based on numerical weather prediction, comprising: The data acquisition module is used to acquire multi-element data of the numerical weather prediction spatial grid for the target area; The feature construction module is used to construct features from multi-factor data to obtain time-domain features and spatial-domain features. The feature fusion module concatenates the time-domain features and spatial-domain features to obtain spatiotemporal features; The training module is used to train the preset point prediction model and the preset quantile regression model to obtain the trained point prediction model and quantile regression model. The prediction module is used to input spatiotemporal features into the trained point prediction model and quantile regression model, and output the prediction results of new energy power generation.

[0025] Compared with the prior art, the beneficial effects of the technical solution of the present invention are: This invention provides a method and system for predicting new energy power generation based on numerical weather prediction. First, multi-element data of the spatial grid of numerical weather prediction for the target area are acquired, and features are constructed to obtain time-domain and spatial-domain features. The time-domain and spatial-domain features are then concatenated to obtain spatiotemporal features. Pre-set point prediction models and quantile regression models are trained to obtain trained point prediction models and quantile regression models. Finally, the spatiotemporal features are input into the trained point prediction models and quantile regression models to output the new energy power generation prediction results. This invention improves the accuracy of new energy power generation prediction by extracting the spatiotemporal features of multi-element data from the spatial grid of numerical weather prediction. Attached Figure Description

[0026] Figure 1 A flowchart illustrating the new energy power generation prediction method based on numerical weather forecasting proposed in this embodiment of the invention; Figure 2 A structural diagram of the new energy power generation prediction device based on numerical weather forecasting proposed in this embodiment of the invention; Figure 3 This diagram illustrates the structure of the new energy power generation prediction system based on numerical weather forecasting proposed in this embodiment of the invention. Detailed Implementation

[0027] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent. To better illustrate this embodiment, some parts of the accompanying drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions; It is understandable to those skilled in the art that some well-known details may be omitted from the accompanying drawings.

[0028] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0029] The positional relationships depicted in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent. Example 1 This embodiment proposes a new energy power generation forecasting method based on numerical weather prediction, such as... Figure 1 The flowchart shown illustrates the method, and the method proposed in this embodiment generally includes the following steps: S1: Acquire multi-element data of the numerical weather prediction spatial grid for the target area; S2: Construct features from multi-factor data to obtain time-domain features and spatial-domain features; S3: Combine the time-domain features and spatial-domain features to obtain spatiotemporal features; S4: Train the preset point prediction model and the preset quantile regression model to obtain the trained point prediction model and quantile regression model. S5: Input the spatiotemporal features into the trained point prediction model and quantile regression model, and output the prediction results of new energy power generation.

[0030] In this embodiment, the multi-element data of the numerical weather prediction (NWP) spatial grid for the target area includes: photovoltaic power generation data and wind power generation data; Photovoltaic power generation data is surface shortwave irradiance. Temperature at a height of 2 meters Low cloud cover Medium cloud cover High cloud cover Total cloud cover ; Wind power generation data consists of the east-west wind speed components of the multi-mode layer. East-west wind speed component Wind speed modulus and wind direction .

[0031] In this embodiment, the multi-element data of the numerical weather forecast spatial grid of the target area comes from a high-resolution mesoscale meteorological model. A 13×13 equidistant grid with a spatial resolution of 4km and a temporal resolution of 1 hour is used to cover an area of ​​approximately 2400km² to capture local and regional weather dynamics.

[0032] In this embodiment, step S1 further includes: Acquire historical power data for the target area, and align the timestamps of the historical power data for the target area with the multi-element data of the numerical weather prediction spatial grid; fill in missing values ​​in the historical power data, and identify and remove outliers in the historical power data. Based on historical surface shortwave irradiance data and historical power data, a clear sky baseline curve is generated. Based on the clear sky baseline curve, the historical power data and surface shortwave irradiance data of the target area are normalized to obtain the normalized power index characterizing weather fluctuations.

[0033] For example, the missing rate of historical power data is controlled to be ≤5%, and missing values ​​are filled by interpolation or trend filling of data from multiple branches at the same site to ensure data integrity.

[0034] Furthermore, the process of obtaining the time-domain features is as follows: Key meteorological variables were selected from multi-element data of numerical weather prediction spatial grids based on the type of new energy source. x ; For photovoltaic power generation, key meteorological variables The preferred parameters are surface shortwave irradiance and total cloud cover. For wind power generation, key meteorological variables... The preferred value is the wind speed modulus at the hub height.

[0035] Set up a sliding window to extract key meteorological variables. x Calculate key meteorological variables x The variance of the time series within the sliding window is expressed as:

[0036] In the formula, express The time series variance at time t. t Let k represent the current time, and k represent the prediction lead time. Indicates the target prediction time. Indicates the length of the sliding window. i This represents the time index of the time window. This represents the arithmetic mean of the values ​​of numerical weather prediction variables. This represents the value of the numerical weather forecast variable at time i. For example, obtain The time series variance of the key meteorological variable x within the solar photovoltaic (PV) system. Time series variance can quantitatively characterize the short-term volatility of NWP variables, such as high [variable name] in solar photovoltaic scenarios. The sudden change in radiation corresponds to cloudy weather.

[0037] Based on key meteorological variables x，Construct the time-lag characteristics of historical values ​​several times before the prediction time and the leading characteristics of predicted values ​​several times after the prediction time. By introducing temporal context information before and after the prediction time, the temporal autocorrelation of meteorological data is extracted. By explicitly encoding future trends (leading terms) and historical inertial information (time lag terms) as input features, the continuity and gradient changes of the NWP time series are captured, and time-domain features are constructed.

[0038] For the same prediction time Several key meteorological variables were obtained, and the multi-time fusion feature was calculated. The expression is as follows:

[0039] In the formula, Indicates multi-temporal fusion features, Indicates weight, Indicates the dth time Forecast value for the time; For example, four key meteorological variables were obtained. .

[0040] Weight Solving using Bayesian optimization, the objective is to maximize... ( (Based on historical power) to reduce the randomness error of a single NWP forecast.

[0041] Based on the time series variance, time lag features, lead features, and multi-time fusion features, time domain features are obtained.

[0042] Furthermore, the process of obtaining spatial domain features is as follows: Based on multi-factor data, at the prediction time Calculate the spatial standard deviation for each spatial grid, expressed as:

[0043] In the formula, Indicates spatial standard deviation. This represents the numerical weather forecast variable values ​​for i spatial grids. Indicates the number of spatial grids. express The mean value of all spatial grid numerical weather forecast variables at any given time can reflect the spatial uniformity of weather distribution; the higher the value, the stronger the spatial heterogeneity.

[0044] Centered on the target site in the target area, the spatial grid is divided into L layers. The smooth mean of all grids in each layer is calculated. The smooth mean of all layers is weighted and fused to obtain the smooth feature. For example, the spatial grid is divided into 6 layers.

[0045] The expression for calculating the smoothed mean is as follows:

[0046] In the formula, express The smoothed mean at time t, Represents the set of grids in the i-th layer; express At time j, the numerical weather forecast variable value of the j-th grid; The calculation expression for the smoothing feature is as follows:

[0047] In the formula, express Smoothness characteristics of time, Indicates the weight of the i-th layer; Solving for weights using Bayesian optimization algorithm The objective function is set as: maximizing the constructed spatial smoothness features. Compared with historical actual power data The Pearson correlation coefficient between them. Through this optimization process, the model can automatically identify the spatial range that has the most significant impact on power generation, thereby effectively balancing local and regional weather information.

[0048] Based on multi-element data from the spatial grid of numerical weather prediction, historical event sequence data of all grids are extracted to construct a historical spatiotemporal data matrix. ;in, Number of time samples (rows). The number of spatial grid points (rows) is given; the historical spatiotemporal data matrix is ​​standardized to obtain the standard historical spatiotemporal data matrix; singular value decomposition is performed on the standard historical spatiotemporal data matrix; the first r principal components are selected; the original high-dimensional spatial data is projected onto the low-dimensional principal component space to obtain the dimensionality-reduced principal component features; The expression for the principal component features is:

[0049] In the formula, Z Indicates principal component characteristics, Let represent the matrix of the first r right singular vectors.

[0050] Historical spatiotemporal data matrix After standardization and singular value decomposition (SVD), the expression is:

[0051] In the formula, U represents the left singular vector matrix and V represents the right singular vector matrix.

[0052] Take before The right singular vectors corresponding to the maximal singular values ​​are used as principal components, so that the cumulative variance contribution rate reaches the threshold.

[0053] For example, threshold 95%.

[0054] Principal component features are performed separately according to variable type, which can reduce feature dimensionality while preserving key spatial modes, such as in photovoltaics. More than 90% of the variance can be explained by just one principal component.

[0055] Based on the spatial standard deviation, smoothing features, and principal component features, spatial domain features are obtained.

[0056] Furthermore, the process of step S3 is as follows: By concatenating the time-domain features and spatial-domain features, a set of candidate features is obtained. Calculate the Pearson correlation coefficients of all features in the candidate feature set and the acquired historical power data, and delete features whose Pearson correlation coefficients are lower than a preset threshold to obtain the first feature subset. Calculate the variance inflation factor among all features in the first feature subset, delete features whose variance inflation factor value is higher than a preset factor threshold, and obtain the second feature subset; Based on the preset gradient boosting tree base learner, the importance scores of the features in the second feature subset are calculated and sorted from high to low according to the importance scores. Features with importance scores lower than the preset importance threshold are deleted to obtain the spatiotemporal features.

[0057] In this embodiment, step S4 specifically includes: Based on time-domain and spatial-domain features, a candidate model library containing N models is constructed. The database includes: a time feature model with input features of time lag and lead features, a time enhancement model with input features of time-series fusion features and time series variance, a spatial feature model with input features of spatial standard deviation and smoothing features, a spatial dimensionality reduction model with input features of principal component features, and a fusion model with input features of spatial domain and time domain features. Each model in the training candidate model library has a preset point prediction model and a preset quantile regression model; Calculate the comprehensive index of each model in the candidate model library, and select the model with the smallest comprehensive index as the optimal model; The preset point prediction model and preset quantile regression model of the optimal model are used as the trained point prediction model and quantile regression model.

[0058] For example, a preferred fusion model can achieve higher prediction accuracy through the synergistic effect of spatiotemporal features.

[0059] A gradient boosting tree (GBT) model is used to construct point prediction and quantile regression models. The GBT model consists of cascaded processing layers of several base learners, preferably Classification and Regression Trees (CART). The base learners are connected sequentially using a boosting method. During training, the model is not trained in parallel, but rather sequentially. Tree of return Before fitting The residuals or negative gradient directions of the loss function of the tree ensemble model are used to gradually correct the prediction error. The process of training the preset point prediction model and preset quantile regression model for each model in the candidate model library is as follows: During the training of the point prediction model, a point prediction loss function is constructed. When the loss function converges, the trained point prediction model is obtained. The point prediction loss function is either a squared loss function or an absolute bias loss function. The expression for the squared loss function is:

[0060] In the formula, Let M represent the squared loss function, M represent the total number of training samples, and m represent the index of the training sample. This represents the true value of the m-th sample. This represents the model prediction value for the m-th sample;

[0061] In the formula, Let M represent the absolute bias loss function, M represent the total number of training samples, and m represent the index of the training sample. This represents the true value of the m-th sample. This represents the model prediction value for the m-th sample; During training, the quantile regression model is updated iteratively in an additive manner, and the expression for the update process is:

[0062]

[0063] In the formula, Indicates the first A tree of return, Indicates the learning rate. The tree hyperparameters are represented, and the point prediction accuracy is improved by progressively optimizing the loss function.

[0064] For a pre-set set of target quantiles The gradient boosting tree algorithm for quantile regression is used to train Q independent models to form a quantile regression model. For each quantile, a quantile loss function is constructed during the training of the quantile regression model. When the loss function converges, the trained quantile regression model is obtained. The expression for the quantile loss function is:

[0065] In the formula, Let M represent the quantile loss function, M represent the total number of training samples, and m represent the index of the training sample. This represents the true value of the m-th sample. Indicates fractional place The predicted value, This represents the quantile loss kernel function.

[0066] The expression for the quantile loss kernel function is:

[0067] In the formula, This represents the predicted residual.

[0068] Output quantile array It is used to quantify and predict uncertainty.

[0069] The training set duration is set to at least 12 months, and the validation set (test window) duration is set to 5 months. The model is evaluated by sliding the window forward month by month.

[0070] In this embodiment, the calculation expression for the comprehensive index is:

[0071] In the formula, express, Indicates the mean absolute error. This represents the root mean square error. This represents the root mean square error. Indicates the weighting coefficient. Indicates the weighting coefficient. This represents the weighting coefficient.

[0072] The expression for the mean absolute error is:

[0073] In the formula, This represents the actual power generation at time t. This represents the predicted value at time t. This indicates the rated installed capacity of the power plant; The expression for the root mean square error is:

[0074] In the formula, This represents the actual power generation at time t. This represents the predicted value at time t. This indicates the rated installed capacity of the power station.

[0075] The expression for the root mean square error is:

[0076] In the formula, The cumulative distribution function predicted at time i, where I represents the indicator function.

[0077] In this embodiment, multi-fold cross-validation is used to ensure the generalization ability of the model in different scenarios.

[0078] The relative improvement rate is used to quantify and evaluate the predictive performance improvement effect of the method of the present invention. The calculation expression of the relative improvement rate is as follows:

[0079] In the formula, express, This represents the monthly average performance metric of the baseline model. This represents the monthly average performance metric of the model being evaluated.

[0080] A baseline model is defined, whose input features include only numerical weather forecast variables and basic time characteristics of the grid center point where the target power station is located. Using any model from this invention as the evaluation model, the monthly average performance index of the baseline model and the evaluation model on the same validation set is calculated. The relative improvement rate of the evaluation model relative to the baseline model is calculated.

[0081] A positive relative improvement rate indicates that the performance of the model under evaluation is better than that of the baseline model. In the 24-hour prediction time domain, for photovoltaic and wind power scenarios, the average relative improvement rate of the model under evaluation on normalized mean absolute error and normalized continuous graded probability score is not less than 10%, and the improvement is even higher in some months.

[0082] The present invention proposes a systematic formulaic construction for spatiotemporal feature engineering. , , , This approach utilizes PCA to quantitatively characterize weather variability and spatial heterogeneity, overcoming the limitations of traditional single-point NWP forecasting. It employs interpretable and scalable GBT / quantile regression to unify coverage of point and probabilistic forecasts, with measurably optimized metrics (MAE, RMSE, CRPS). Hierarchical smoothing complements PCA, balancing local and large-scale modes, maintaining robustness in small samples and long time domains, and offering high computational efficiency and engineering feasibility.

[0083] Example 2 This invention also provides a new energy power generation prediction device based on numerical weather forecasting, the structure of which is shown in the figure below. Figure 2 The following are included: The data acquisition unit is used to acquire multi-element data of the numerical weather prediction spatial grid of the target area and historical power data of the target area; The processor unit is communicatively connected to the data acquisition unit and is used to run a pre-stored computer program to execute the new energy power generation prediction method based on numerical weather forecasting, and generate power generation point prediction values ​​and probability prediction intervals for future periods. The display unit, connected to the processor unit, is used to display the point prediction value, probability prediction interval, and related prediction curve; The power supply unit is used to supply power to the data acquisition unit, the processor unit, and the display unit.

[0084] In this embodiment, the processor unit is a Raspberry Pi 4B or an industrial computer. The pre-deployed prediction models on the processor are a point prediction model and a quantile regression model based on Gradient Boosting Tree (GBT). The GBT model uses a regression tree as the base learner and optimizes the loss function through additive iterative updates. The quantile regression model is designed for... A total of 19 quantiles were trained separately, and the model hyperparameters (such as maximum tree depth, learning rate, and number of iterations) were solved using a Bayesian optimization algorithm to improve the model's generalization ability.

[0085] The data acquisition unit includes an NWP grid data interface and a historical power data interface, wherein the NWP grid data interface is used to acquire data from the target area. Multi-element meteorological forecast data with equidistant grids; historical power data interface for acquiring historical output data of wind / solar power plants; optional access to local environmental sensors for calibration and feature enhancement. The NWP multi-element data is categorized by energy type: photovoltaic data includes at least surface shortwave irradiance. 2m temperature Low / Medium / High / Total Cloud Cover Wind power should include wind vectors from at least a multi-mode layer. , Wind speed modulus With wind direction The original NWP variable dimensions reach 1000+.

[0086] The processor unit receives the latest NWP multi-temporal and spatial grid data, calculates features in real time, and generates point predictions for the next 1–72 hours. With quantiles The predictive performance continues to improve positively as time increases.

[0087] For example, the time resolution is preferably 1 hour, and the prediction time domain is 0–72 hours.

[0088] Display unit outputs prediction curve and confidence interval It also provides API integration with scheduling and trading systems to support power grid operation decisions.

[0089] Example 3 This invention also provides a new energy power generation prediction system based on numerical weather forecasting, the system structure diagram of which is shown below. Figure 3 As shown, it includes: The data acquisition module is used to acquire multi-element data of the numerical weather prediction spatial grid for the target area; The feature construction module is used to construct features from multi-factor data to obtain time-domain features and spatial-domain features. The feature fusion module concatenates the time-domain features and spatial-domain features to obtain spatiotemporal features; The training module is used to train the preset point prediction model and the preset quantile regression model to obtain the trained point prediction model and quantile regression model. The prediction module is used to input spatiotemporal features into the trained point prediction model and quantile regression model, and output the prediction results of new energy power generation.

[0090] The various embodiments in this invention are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The device embodiments described above are merely exemplary. The modules described as separate components may or may not be physically separate. When implementing the present invention, the functions of each module can be implemented in one or more software and / or hardware. Alternatively, some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0091] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for predicting new energy power generation based on numerical weather forecasting, characterized in that, Includes the following steps: S1: Acquire multi-element data of the numerical weather prediction spatial grid for the target area; S2: Construct features from multi-factor data to obtain time-domain features and spatial-domain features; S3: Combine the time-domain features and spatial-domain features to obtain spatiotemporal features; S4: Train the preset point prediction model and the preset quantile regression model to obtain the trained point prediction model and quantile regression model. S5: Input the spatiotemporal features into the trained point prediction model and quantile regression model, and output the prediction results of new energy power generation.

2. The new energy power generation prediction method based on numerical weather prediction according to claim 1, characterized in that, The multi-element data of the numerical weather forecast spatial grid for the target area includes: photovoltaic power generation data and wind power generation data; Photovoltaic power generation data is surface shortwave irradiance. Temperature at a height of 2 meters Low cloud cover Medium cloud cover High cloud cover Total cloud cover ; Wind power generation data consists of the east-west wind speed components of the multi-mode layer. East-west wind speed component Wind speed modulus and wind direction .

3. The new energy power generation forecasting method based on numerical weather prediction according to claim 2, characterized in that, Step S1 also includes: Acquire historical power data for the target area, and align the timestamps of the historical power data for the target area with the multi-element data of the numerical weather prediction spatial grid; fill in missing values ​​in the historical power data, and identify and remove outliers in the historical power data. Based on historical surface shortwave irradiance data and historical power data, a clear sky baseline curve is generated. Based on the clear sky baseline curve, the historical power data and surface shortwave irradiance data of the target area are normalized to obtain the normalized power index characterizing weather fluctuations.

4. The new energy power generation forecasting method based on numerical weather prediction according to claim 1, characterized in that, The process of obtaining time-domain features is as follows: Key meteorological variables were selected from multi-element data of numerical weather prediction spatial grids based on the type of new energy source. x ; Set up a sliding window to extract key meteorological variables. x Calculate key meteorological variables x The variance of the time series within the sliding window is expressed as: In the formula, express The time series variance at time t. t Let k represent the current time, and k represent the prediction lead time. Indicates the target prediction time. Indicates the length of the sliding window. i This represents the time index of the time window. This represents the arithmetic mean of the values ​​of numerical weather prediction variables. This represents the value of the numerical weather forecast variable at time i. Based on key meteorological variables x， Construct time-lag features that include historical values ​​several moments before the prediction time and advance features that include predicted values ​​several moments after the prediction time; For the same forecast time, several key meteorological variables are obtained, and the multi-time fusion feature is calculated. The expression is as follows: In the formula, Indicates multi-temporal fusion features, Indicates weight, Indicates the dth time Forecast value for the time; Based on the time series variance, time lag features, lead features, and multi-time fusion features, time domain features are obtained.

5. The new energy power generation prediction method based on numerical weather prediction according to claim 1, characterized in that, The process of obtaining spatial domain features is as follows: Based on multi-factor data, at the prediction time Calculate the spatial standard deviation for each spatial grid, expressed as: In the formula, Indicates spatial standard deviation. This represents the numerical weather forecast variable values ​​for i spatial grids. Indicates the number of spatial grids. express The mean value of all spatial grid numerical weather forecast variables at any given time; Centered on the target site in the target area, the spatial grid is divided into L layers. The smooth mean of all grids in each layer is calculated. The smooth mean of all layers is weighted and fused to obtain the smooth feature. The expression for calculating the smoothed mean is as follows: In the formula, express The smoothed mean at time t, Represents the set of grids in the i-th layer; express At time j, the numerical weather forecast variable value of the j-th grid; The calculation expression for the smoothing feature is as follows: In the formula, express Smoothness characteristics of time, Indicates the weight of the i-th layer; Based on multi-element data from the spatial grid of numerical weather prediction, historical event sequence data of all grids are extracted to construct a historical spatiotemporal data matrix. ;in, For the number of time samples, The number of spatial grid points is given; the historical spatiotemporal data matrix is ​​standardized to obtain a standard historical spatiotemporal data matrix; singular value decomposition is performed on the standard historical spatiotemporal data matrix; the first r principal components are selected; the original high-dimensional spatial data is projected onto the low-dimensional principal component space to obtain the dimensionality-reduced principal component features. Based on the spatial standard deviation, smoothing features, and principal component features, spatial domain features are obtained.

6. The new energy power generation prediction method based on numerical weather prediction according to claim 1, characterized in that, The process of step S3 is as follows: By concatenating the time-domain features and spatial-domain features, a set of candidate features is obtained. Calculate the Pearson correlation coefficients of all features in the candidate feature set and the acquired historical power data, and delete features whose Pearson correlation coefficients are lower than a preset threshold to obtain the first feature subset. Calculate the variance inflation factor among all features in the first feature subset, delete features whose variance inflation factor value is higher than a preset factor threshold, and obtain the second feature subset; Based on the preset gradient boosting tree base learner, the importance scores of the features in the second feature subset are calculated and sorted from high to low according to the importance scores. Features with importance scores lower than the preset importance threshold are deleted to obtain the spatiotemporal features.

7. The new energy power generation prediction method based on numerical weather prediction according to claim 1, characterized in that, Step S4 specifically involves: Based on time-domain and spatial-domain features, a candidate model library containing N models is constructed. The database includes: a time feature model with input features of time lag and lead features, a time enhancement model with input features of time-series fusion features and time series variance, a spatial feature model with input features of spatial standard deviation and smoothing features, a spatial dimensionality reduction model with input features of principal component features, and a fusion model with input features of spatial domain and time domain features. Each model in the training candidate model library has a preset point prediction model and a preset quantile regression model; Calculate the comprehensive index of each model in the candidate model library, and select the model with the smallest comprehensive index as the optimal model; The preset point prediction model and preset quantile regression model of the optimal model are used as the trained point prediction model and quantile regression model.

8. A new energy power generation forecasting method based on numerical weather prediction according to claim 7, characterized in that, The process of training the preset point prediction model and preset quantile regression model for each model in the candidate model library is as follows: During the training of the point prediction model, a point prediction loss function is constructed. When the loss function converges, the trained point prediction model is obtained. The point prediction loss function is either a squared loss function or an absolute bias loss function. The expression for the squared loss function is: In the formula, Let M represent the squared loss function, M represent the total number of training samples, and m represent the index of the training sample. This represents the true value of the m-th sample. This represents the model prediction value for the m-th sample; In the formula, Let M represent the absolute bias loss function, M represent the total number of training samples, and m represent the index of the training sample. This represents the true value of the m-th sample. This represents the model prediction value for the m-th sample; During the training of the quantile regression model, a quantile loss function is constructed. When the loss function converges, the trained quantile regression model is obtained. The expression for the quantile loss function is: In the formula, Let M represent the quantile loss function, M represent the total number of training samples, and m represent the index of the training sample. This represents the true value of the m-th sample. Indicates fractional place The predicted value, This represents the quantile loss kernel function.

9. A new energy power generation forecasting method based on numerical weather prediction according to claim 7, characterized in that, The formula for calculating the comprehensive index is as follows: In the formula, express, Indicates the mean absolute error. This represents the root mean square error. This represents the root mean square error. Indicates the weighting coefficient. Indicates the weighting coefficient. This represents the weighting coefficient.

10. A new energy power generation forecasting system based on numerical weather prediction, used to implement the new energy power generation forecasting method based on numerical weather prediction as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire multi-element data of the numerical weather prediction spatial grid for the target area; The feature construction module is used to construct features from multi-factor data to obtain time-domain features and spatial-domain features. The feature fusion module concatenates the time-domain features and spatial-domain features to obtain spatiotemporal features; The training module is used to train the preset point prediction model and the preset quantile regression model to obtain the trained point prediction model and quantile regression model. The prediction module is used to input spatiotemporal features into the trained point prediction model and quantile regression model, and output the prediction results of new energy power generation.

Images