A photovoltaic and wind power generation capacity prediction method and system for a multi-energy complementary system and a medium

Abstract

This invention discloses a weighted integration method, system, and medium for predicting photovoltaic and wind power generation in multi-energy complementary systems, belonging to the field of new energy power generation prediction technology. The invention generates photovoltaic weather condition labels and wind power wind condition labels through data preprocessing and operating condition identification. For the preprocessed dataset, multiple training subsets with data diversity are generated, and at least two different base models are trained on each subset to form a set of structurally diverse base models. Error indices of each base model are calculated according to operating condition categories on the validation set, and adaptive fusion weights are constructed. After obtaining the point prediction results, a prediction interval or budget parameter set is used for robust optimization decision-making. This invention improves prediction accuracy under extreme operating conditions, achieves adaptive weighted fusion of operating conditions, outputs an uncertainty set that can directly support robust optimization, and has a standardized engineering interface, facilitating rapid deployment and application in multi-energy complementary systems.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation prediction technology, specifically to a method, system, and medium for predicting photovoltaic and wind power generation for multi-energy complementary systems. Background Technology

[0002] Due to the clean and low-carbon advantages of renewable energy, its share in the power system and integrated energy system is constantly increasing, but it also brings significant power output fluctuations and uncertainties: photovoltaic power output is affected by factors such as solar irradiance, cloud cover changes, and temperature, and there are rapid ramp-up and short-term sudden changes; wind power output is affected by wind speed, wind direction, turbulence intensity, air density, etc., and has obvious randomness and time-varying characteristics.

[0003] In multi-energy complementary systems, photovoltaic and wind power are typically coupled with energy storage, gas turbines, thermal power, pumped storage, and grid power purchase and sale. System day-ahead planning (e.g., 24-hour or 1-hour resolution) requires advance determination of unit output, energy storage charging and discharging strategies, and power purchase curves. Therefore, power generation forecasting is central to multi-energy complementary calculations: forecasting errors can lead to increased power balance deviations, increased reserve demand, failure of energy storage strategies, increased wind and solar curtailment, and even constraint violations and decreased economic efficiency. Constructing high-precision, highly stable day-ahead forecasting models and providing reliable uncertainty descriptions for optimization decisions are crucial foundations for the safe and economical operation of multi-energy complementary systems.

[0004] Existing photovoltaic and wind power forecasting methods are mainly divided into three categories: methods based on physical mechanisms, methods based on statistical learning, and methods based on machine learning / deep learning. Methods based on physical mechanisms rely on models such as irradiance transfer or wind energy conversion, are greatly affected by parameter identification and on-site conditions, and are sensitive to local weather changes. Statistical learning methods (such as ARIMA) are applicable to linear and weakly nonlinear sequences, but are insufficient for responding to strong nonlinearities and abrupt changes under complex operating conditions. Machine learning / deep learning methods (such as neural networks, support vector machines, tree models, etc.) can fit complex nonlinearities, but still have the following prominent problems in practical engineering applications: (1) The model has insufficient generalization ability and poor response to extreme working conditions. The data distribution varies significantly under different weather types (sunny, cloudy, rainy) or different wind conditions (low wind, medium wind, strong wind). Furthermore, the proportion of common operating conditions in real stations is too high, while the samples of extreme or low-probability operating conditions (such as sudden cloud formations or gusts) are scarce. This causes a single model to be prone to bias towards the mainstream distribution, and the error surges under extreme operating conditions.

[0005] (2) The multi-model fusion strategy is crude and lacks adaptability to operating conditions. Existing fusion methods mostly employ simple averaging or fixed weights, with weights that do not change with operating conditions, making it difficult to consider the optimal model combination under different operating conditions. Furthermore, they lack a systematic design that incorporates data diversity, resulting in limited fusion benefits.

[0006] (3) Only outputting point predictions cannot directly support robust optimization. Many methods only provide point predictions and cannot quantify the uncertainty range of prediction errors. Day-ahead decision-making for multi-energy complementary systems requires prediction intervals or budget parameters to construct robust constraints; otherwise, plans may become infeasible, reserves may be insufficient, or wind and solar power curtailment may increase.

[0007] (4) Insufficient engineering adaptability and non-standard interfaces. Predictive models are difficult to integrate with energy management systems (EMS) and scheduling optimization modules in a standardized manner, and the output formats (point values, intervals, budget parameters) are inconsistent, affecting the actual operation closed loop.

[0008] In summary, there is an urgent need for a power generation prediction method that can simultaneously cover photovoltaic and wind power, has adaptive fusion capability under operating conditions, outputs an uncertainty set, and is easy to integrate into engineering projects. Summary of the Invention

[0009] The purpose of this invention is to solve the problems mentioned in the background section above, and this invention provides the following technical solution: A method for predicting photovoltaic and wind power generation in a multi-energy complementary system includes the following steps: S1. Obtain historical photovoltaic power output data, historical wind power output data and corresponding weather forecast data, and preprocess the historical photovoltaic power output data, the historical wind power output data and the corresponding weather forecast data to obtain a preprocessed historical dataset; S2. Based on the preprocessed historical dataset, identify the weather conditions for photovoltaic power generation and the wind conditions for wind power generation, and generate corresponding condition labels. S3. For the preprocessed historical dataset, use one or more of the following methods: stratified sampling, resampling, or time slicing to generate multiple training subsets with data diversity. S4. On any of the training subsets, train at least two base models with different structures to form a base model set, and obtain the prediction output of each base model. S5. Divide the preprocessed dataset into a training set and a validation set, wherein the training set is used to generate multiple training subsets in step S3; on the validation set, calculate the error index of each base model according to the working condition category, and construct working condition adaptive fusion weights based on the error index. S6. Based on the current working condition label, call the corresponding fusion weight to perform weighted fusion on the prediction output of each base model to obtain the final point prediction result; S7. Based on the prediction divergence degree and / or historical prediction residual distribution of the base model set, determine the uncertainty scale and construct an uncertainty set for robust optimization, wherein the uncertainty set is a prediction interval set or a budget parameter set. S8. Output the point prediction results and the uncertainty set for robust optimization decision-making of the multi-energy complementary system.

[0010] More preferably, the weather conditions for the photovoltaic system include at least three categories: sunny, cloudy, and overcast / rainy, which are classified based on irradiance fluctuations, clear sky index, or cloud cover. The wind conditions for wind power include at least three categories: low wind, medium wind, and strong wind, or two categories: stable and fluctuating, based on wind speed range, wind speed fluctuation rate, or turbulence intensity.

[0011] More preferably, in step S3, the specific methods for generating K training subsets with data diversity include a combination of at least two of the following: Stratified sampling: Stratified sampling is carried out according to working condition labels, seasons, months, or weekdays / weekends to ensure coverage of each working condition; Resampling: The Bootstrap resampling method is used to construct training subsets with different sample distributions; Time slicing: Divide the original data into multiple training subsets based on time windows to form different time periods.

[0012] More preferably, in step S5, the fusion weights for constructing the adaptive operating condition are calculated using an inverse error normalization method:

[0013] in, For the first On the training subset, the first Individual basic models under working conditions The weights below, For the base model under the working condition The error indicators include at least one of mean absolute error, root mean square error, or mean absolute percentage error. is the regularization coefficient.

[0014] More preferably, the base model in step S4 includes, but is not limited to, at least two of the following: Generalized Regression Neural Network (GRNN), Extreme Learning Machine (ELM), Elman Neural Network (ElmanNN), Long Short-Term Memory Network (LSTM), or Extreme Gradient Boosting (XGBoost) model.

[0015] More preferably, in step S7: When the set of uncertainties is a set of prediction intervals, it is represented as:

[0016] in, and These are the lower and upper bounds of the forecast, respectively. When the set of uncertainties is a set of budget parameters, it is represented as:

[0017] And increase budget constraints:

[0018] in, Indicates the current moment. Indicates the prediction step size. , To predict the total number of time periods, This represents the actual amount of electricity generated. For point prediction results, To predict bias, The uncertainty scale is determined by the standard deviation of the forecast or the quantile of the historical residual distribution based on the set of base models. For budget parameters, subscript Represents photovoltaic or wind power .

[0019] More preferably, in step S1, the meteorological forecast data for photovoltaic power includes at least irradiance characteristics, and the meteorological forecast data for wind power includes at least wind speed characteristics. In step S8, the point prediction results and uncertainty set are output at a time scale of 24-hour and 1-hour resolution.

[0020] A weighted integrated system for predicting photovoltaic and wind power generation in multi-energy complementary systems, used to implement the photovoltaic and wind power generation prediction method for multi-energy complementary systems as described in any of the preceding claims, comprising: Data processing module: used to perform data acquisition, alignment, cleaning, normalization, and feature construction; Operating condition identification module: used to generate weather operating condition labels for photovoltaic power and wind operating condition labels for wind power; Sub-training set generation module: used to generate multiple training subsets with data diversity; Base Model Training and Inference Module: Used to train multiple base models with different structures and obtain their prediction outputs; Weight learning and fusion module: used to calculate the adaptive weights for working conditions and output the weighted fusion point prediction results; Uncertainty assessment module: used to output forecast intervals or budget set parameters; Interface module: Used to standardize and output the point prediction results and uncertainty set parameters to the robust optimization module or energy management system of the multi-energy complementary system.

[0021] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the photovoltaic and wind power generation prediction method for multi-energy complementary systems as described in any of the preceding claims.

[0022] Compared with existing technologies, the weighted integration method, system, and medium for predicting photovoltaic and wind power generation in multi-energy complementary systems provided by this invention can bring the following significant benefits: (1) Improve the model's generalization ability and enhance the prediction accuracy for extreme working conditions. This invention enhances data diversity (stratified sampling, resampling, and time slicing) to cover multiple operating conditions, and combines a set of structurally diverse base models to effectively reduce performance degradation caused by data distribution drift and extreme sample scarcity, significantly reducing prediction errors under low-probability operating conditions.

[0023] (2) Achieve adaptive weighted fusion of working conditions and overcome the defects of fixed weights. This invention calculates the error index of each base model according to the working condition category and constructs a fusion weight that dynamically changes with weather / wind conditions, so that the optimal model combination can be automatically selected under different working conditions, resulting in a more stable fusion output and higher accuracy.

[0024] (3) Output the uncertainty set to directly support robust optimization decision-making. This invention not only outputs point predictions, but also constructs prediction intervals or budget parameter sets based on model divergence and historical residual distribution. These can be directly embedded into the robust optimization model of a multi-energy complementary system to build power balance constraints and reserve constraints, reduce scheduling deviation risks, and reduce wind and solar curtailment and reserve pressure.

[0025] (4) Strong engineering adaptability and standardized interfaces This invention designs a unified output interface (point prediction + uncertainty set), which can be easily connected to EMS and scheduling optimization modules, and supports standardized input with 24-hour and 1-hour resolutions, making it easy to implement quickly at different sites. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of photovoltaic operating condition identification using the method of the present invention; Figure 3 This is a schematic diagram of wind power operating condition identification using the method of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] The preferred embodiments of the present invention will now be described in further detail with reference to the accompanying drawings.

[0029] This embodiment provides a method for predicting photovoltaic and wind power generation in multi-energy complementary systems, such as... Figure 1 The method flowchart of this invention illustrates the principle of weighted integration. The prediction method in this embodiment includes the following steps: S1. Data Acquisition and Preprocessing: Acquire historical photovoltaic power output data, historical wind power output data, and corresponding weather forecast data, and preprocess the data to obtain a preprocessed dataset. The weather forecast data for photovoltaic power shall at least include irradiance characteristics, and the weather forecast data for wind power shall at least include wind speed characteristics.

[0030] S2. Operating Condition Identification: Based on the preprocessed data, identify the weather conditions for photovoltaic power and the wind conditions for wind power, and generate corresponding operating condition labels. The weather conditions for photovoltaic power include at least three categories: sunny, cloudy, and overcast / rainy, classified based on irradiance fluctuations, clear sky index, or cloud cover; the wind conditions for wind power include at least three categories: low wind, medium wind, and strong wind, or two categories: stable and fluctuating, classified based on wind speed range, wind speed fluctuation rate, or turbulence intensity.

[0031] S3. Data Diversity Enhancement: For the preprocessed dataset, K training subsets with data diversity are generated through one or more of the following methods: stratified sampling, resampling, or time slicing. The specific methods for generating the K training subsets with data diversity include combinations of at least two of the following: Stratified sampling: Stratified sampling is carried out according to working condition labels, seasons, months, or weekdays / weekends to ensure coverage of each working condition; Resampling: The Bootstrap resampling method is used to construct training subsets with different sample distributions; Time slicing: Divide the original data into multiple training subsets based on time windows to form different time periods.

[0032] S4. Training of Base Models with Diverse Structures: On any of the training subsets, train at least two base models with different structures to form a base model set, and obtain the prediction output of each base model. The base models include, but are not limited to, at least two of the following models: Generalized Regressive Neural Network (GRNN), Extreme Learning Machine (ELM), Elman Neural Network (ElmanNN), Long Short-Term Memory Network (LSTM), or Extreme Gradient Boosting (XGBoost).

[0033] S5. Working Condition Related Error Assessment and Weight Learning: The preprocessed dataset is divided into a training set and a validation set, wherein the training set is used to generate K training subsets in step S3; on the validation set, the error index of each base model is calculated according to the working condition category, and working condition adaptive fusion weights are constructed based on the error indexes. The working condition adaptive fusion weights are calculated using an inverse error normalization method:

[0034] in, For the first On the training subset, the first Individual basic models under working conditions The weights below, For the base model under the working condition The error indicators include at least one of mean absolute error, root mean square error, or mean absolute percentage error. is the regularization coefficient.

[0035] S6. Adaptive weighted fusion point prediction: Based on the current working condition label, the corresponding fusion weight is called to perform weighted fusion on the prediction output of each base model to obtain the final point prediction result.

[0036] S7. Uncertainty Set Construction: Based on the prediction divergence degree and / or historical prediction residual distribution of the base model set, determine the uncertainty scale and construct an uncertainty set for robust optimization, wherein the uncertainty set is a prediction interval set or a budget parameter set.

[0037] When the set of uncertainties is a set of prediction intervals, it is represented as:

[0038] in, and These are the lower and upper bounds of the forecast, respectively. When the set of uncertainties is a set of budget parameters, it is represented as:

[0039] And increase budget constraints:

[0040] in, Indicates the current moment. Indicates the prediction step size. , To predict the total number of time periods, This represents the actual amount of electricity generated. For point prediction results, To predict bias, The uncertainty scale is determined by the standard deviation of the forecast or the quantile of the historical residual distribution based on the set of base models. For budget parameters, subscript Represents photovoltaic or wind power .

[0041] S8. Output and Connection: Output the point prediction results and the uncertainty set for robust optimization decision-making in the multi-energy complementary system. The point prediction results and uncertainty set are output at a time scale of 24-hour and 1-hour resolution.

[0042] This embodiment provides a weighted integration method for predicting photovoltaic and wind power generation in multi-energy complementary systems. This method can be applied to computer equipment (such as servers, industrial control computers, or cloud platforms). By executing the following steps S1 to S8, it can achieve prediction of photovoltaic and wind power generation points and output of uncertainty sets at 24-hour and 1-hour resolutions, which can be used to support the robust optimization scheduling of multi-energy complementary systems.

[0043] Step S1: Data Acquisition and Preprocessing Acquire historical power output data (with a time span of at least one year and a sampling interval of one hour or higher, aggregated into one-hour data) and corresponding weather forecast data for the target photovoltaic power station and wind farm. The photovoltaic-side weather data must at least include total irradiance (or direct / scattered irradiance), and optionally include temperature, humidity, cloud cover, and air pressure. The wind-side weather data must at least include wind speed, and optionally include wind direction, temperature, and air pressure. It also includes time calendar features such as hour (0-23), day type (weekday / weekend), and month.

[0044] Preprocessing includes: Missing value imputation: linear interpolation or forward / backward imputation methods are used; Outlier correction: Identify and correct / remove outliers based on the 3σ principle (three standard deviations) or box plots; Normalization: linearly scales each feature to the interval [0,1] or [-1,1].

[0045] The preprocessed dataset is denoted as . .

[0046] Step S2: Operating Condition Identification Based on the preprocessed data, operating condition labels are generated for each time point, such as... Figure 2 A schematic diagram of photovoltaic operating condition identification method of the present invention. Figure 3 A schematic diagram of wind power operating condition identification using the method of the present invention.

[0047] Photovoltaic operating conditions: Based on irradiance fluctuation rate (the magnitude of irradiance change per unit time), clear sky index (the ratio of actual irradiance to theoretical clear sky irradiance), and cloud cover data, weather is divided into three categories: sunny (small irradiance fluctuation, clear sky index > 0.7), cloudy (0.3 ≤ clear sky index ≤ 0.7, or moderate fluctuation), and overcast / rainy (clear sky index < 0.3, or cloud cover > 80%). Tags are denoted as... .

[0048] Wind power operating conditions: Classified according to wind speed range at 10 meters or hub height into low wind (wind speed <3m / s), medium wind (3m / s ≤ wind speed ≤ 12m / s), and high wind (wind speed > 12m / s); or classified according to wind speed fluctuation rate (1-hour change rate) into stable (fluctuation rate <0.2) and fluctuating (fluctuation rate ≥0.2). Labels are denoted as... or .

[0049] The operating condition label is associated with the data sample at each time point, and is used for subsequent stratified sampling and adaptive weighting.

[0050] Step S3: Enhancing Data Diversity For the preprocessed dataset Generate K training subsets with data diversity. In this embodiment, a combination of the following two methods is used: Stratified sampling: Stratify by operating condition label (sunny / cloudy / rainy, or low / medium / strong wind), season (spring / summer / autumn / winter), and month. Randomly select a certain proportion of samples from each stratum to ensure that the sample proportions of each operating condition and season in each training subset are basically balanced.

[0051] Bootstrap resampling: Based on stratified sampling, each subset is randomly resampled with replacement. The size of the subset is the same as (or slightly smaller than) the original training set, introducing differences in sample distribution.

[0052] For example, set K=10. Each subset contains approximately 80% of the original samples (after stratified sampling to ensure representativeness), and then uses Bootstrap to generate different distributions.

[0053] Step S4: Training the structural diversity base model In each training subset Above, at least two base models with different structures are trained. This embodiment uses the following three models: Extreme Learning Machine (ELM): A single-hidden-layer feedforward neural network that randomly initializes input weights and biases and solves for output weights using least squares. The number of hidden layer neurons is set to 50.

[0054] Long Short-Term Memory Network (LSTM): It consists of two LSTM layers (64 units each) and a fully connected output layer, using the Adam optimizer, with a learning rate of 0.001 and 200 training rounds.

[0055] Extreme Gradient Boosting (XGBoost): Tree model, maximum depth 6, learning rate 0.1, number of subtrees 100.

[0056] Each base model is trained separately for photovoltaic (PV) and wind power. The input feature vectors are the corresponding meteorological and temporal features, and the output is the power generation (or output power) for the next 1 to 24 hours. After training, each base model can output a predicted value for any new input. Let the predicted value of the m-th base model on the k-th training subset at time t+h be denoted as... The predicted value for wind power is ,in , Representative model types (such as ELM, LSTM, XGBoost). .

[0057] This step, combined with the data diversity enhancement in step S3, constitutes the first beneficial effect of this invention: improving the model's generalization ability and enhancing prediction accuracy for extreme conditions. Multiple training subsets covering different conditions are generated through stratified sampling, resampling, or time slicing to avoid model bias towards mainstream samples; simultaneously, at least two different base models (such as ELM, LSTM, and XGBoost) are employed to create structural diversity. The combination of these two approaches effectively reduces performance degradation caused by data distribution drift and the scarcity of extreme samples, significantly reducing prediction errors under low-probability conditions (such as rainy days or windy days).

[0058] Step S5: Assessment of operating condition-related errors and weight learning Preprocessed dataset The dataset is divided into a training set (70%) and a validation set (30%). Note: The training subset in step S3 is generated from the training set, while the validation set is not used for training but only for evaluation and weight learning.

[0059] On the validation set, calculate the error index for each base model (k,m) according to the operating condition category c (e.g., sunny, cloudy, rainy), for example, using the root mean square error (RMSE). Calculate the error index. The calculation formula is:

[0060] in These are common dummy variables. Validation set, where c represents the operating condition. To verify centralized operating conditions The total number of samples, Let be the predicted value for the i-th sample. This represents the actual value of the i-th sample.

[0061] Then, the adaptive weights for the working conditions are calculated based on the inverse error normalization:

[0062] in, For the first On the training subset, the first Individual basic models under working conditions The weights below, For the base model under the working condition The error indicators include at least one of mean absolute error, root mean square error, or mean absolute percentage error. is the regularization coefficient.

[0063] These weights are stored in a weight table, with different weight allocations for each operating condition c. For example, the XGBoost model might have higher weights under sunny conditions, while the LSTM model might have higher weights under cloudy or rainy conditions.

[0064] Step S6: Adaptive weighted fusion point prediction For the current time t to be predicted, first identify its operating condition label according to step S2, and record the photovoltaic operating condition as follows: Wind power operating conditions are recorded as Then call the weight of the corresponding working condition. The predicted outputs of each base model are weighted and fused: Predicted final value for photovoltaic power:

[0065] Wind power final point prediction:

[0066] The final point prediction results are obtained, which have been demonstrated on the validation set to have lower overall error and smaller extreme error than any single base model.

[0067] The aforementioned weighted fusion process demonstrates the second beneficial effect of this invention: achieving adaptive weighted fusion based on operating conditions, overcoming the shortcomings of fixed weights. Step S2 identifies real-time operating conditions (sunny / cloudy / rainy or low / medium / strong winds), and calls the error inverse weights pre-calculated according to the operating conditions in step S5, automatically optimizing the contribution of each base model under different operating conditions—XGBoost weights may be higher on sunny days, while LSTM weights may be higher on rainy days. Compared to simple averaging or fixed-weight fusion methods, the adaptive fusion output of this invention is more stable and has higher accuracy.

[0068] Step S7: Construction of the Uncertainty Set Based on the prediction divergence (i.e., the standard deviation of each base model's output) and the distribution of historical prediction residuals, an uncertainty set for robust optimization is constructed. This embodiment provides two forms, allowing the output to be selected according to the needs of the downstream optimizer: When the set of uncertainties is the set of prediction intervals, it is represented as:

[0069] in, and These are the lower and upper bounds of the forecast, respectively. When the set of uncertainties is the set of budget parameters, it is represented as:

[0070] And increase budget constraints:

[0071] in, Indicates the current moment. Indicates the prediction step size. , To predict the total number of time periods, This represents the actual amount of electricity generated. For point prediction results, To predict bias, The uncertainty scale is determined by the standard deviation of the forecast or the quantile of the historical residual distribution based on the set of base models. For budget parameters, subscript Represents photovoltaic or wind power .

[0072] The uncertainty set (prediction interval or budget parameters) constructed in this step brings a third beneficial effect to this invention: the output uncertainty set directly supports robust optimization decision-making. Compared to existing methods that only output point predictions, this invention provides the possible range of prediction errors (upper and lower limits of the interval) or budget parameters (σ and Γ) with adjustable conservatism. This information can be directly embedded into the robust optimization model of the multi-energy complementary system to construct power balance constraints and reserve constraints, thereby reducing the risk of scheduling deviations and alleviating wind and solar curtailment and reserve pressure.

[0073] Step S8: Output and Interconnection The calculated point prediction results and uncertainty set (interval or budget parameters) are output in a standardized format to the robust optimization module or energy management system (EMS) of the multi-energy complementary system. The output is in JSON or CSV file format and includes fields such as timestamp, predicted value, lower limit, upper limit or standard deviation, and budget parameters.

[0074] The robust optimization module uses these inputs to construct constraints of the following form: For any time When the actual output of photovoltaic power falls within an uncertain set, the system power balance constraint must be satisfied. By introducing dual variables or employing robust equivalent transformations, the optimal day-ahead plan (unit start-up and shutdown, energy storage charging and discharging, power purchase and sale, etc.) under the worst-case scenario can be solved. Compared with point prediction, the prediction and uncertainty output of this invention can significantly reduce the plan infeasibility rate caused by prediction bias.

[0075] The fourth beneficial effect of this invention is reflected in step S8 and the overall system design: strong engineering adaptability and standardized interfaces. Through a unified output interface (point prediction + prediction interval or budget parameters), it can easily connect to EMS and scheduling optimization modules, supporting standardized input at 24-hour and 1-hour resolutions. Simultaneously, all modules in the system (data processing, operating condition identification, sub-training set generation, base model training, weight fusion, uncertainty assessment, and interface) adopt a decoupled design, facilitating the replacement of base model types or weight learning strategies according to actual site conditions, enabling rapid implementation.

[0076] This embodiment verifies the advantages of adaptive fusion weights using real-world data. One year's historical data from a specific wind-solar hybrid power station is selected, and day-ahead forecasts are compared using the following three methods: Method A: Single LSTM model; Method B: Equal-weight fusion of multiple models (simple averaging); Method C: The method of this invention (adaptive weighted fusion based on operating conditions).

[0077] The mean absolute percentage error (MAPE) was statistically analyzed on the validation set (data from the last 3 months) according to the operating conditions. The results are shown in Table 1 below: Table 1. Comparison of prediction errors of different methods under different working conditions.

[0078] It is evident that the method of this invention achieves optimal prediction accuracy under all operating conditions, with particularly significant improvement under extreme conditions (rain, strong winds).

[0079] This embodiment provides a weighted integrated system for predicting photovoltaic and wind power generation in multi-energy complementary systems, used to implement the above method. The system includes the following modules (which can be deployed as software on a server): Data processing module: Performs data acquisition, time alignment, missing value imputation, outlier correction, normalization, and feature construction. Outputs a standardized time-series dataset.

[0080] Operating condition identification module: Generates photovoltaic weather operating condition labels and wind power wind condition operating condition labels based on characteristics such as irradiance and wind speed, and associates them with data samples.

[0081] Sub-training set generation module: Implements stratified sampling, Bootstrap resampling, or time slicing to generate K training subsets with data diversity.

[0082] Base model training and inference module: Parallel training of various base models with different structures such as ELM, LSTM, and XGBoost, and provides an inference interface.

[0083] Weight learning and fusion module: Calculates error indices on the validation set according to working conditions, generates adaptive weights for working conditions, and performs weighted fusion to obtain point prediction results.

[0084] Uncertainty assessment module: Calculates the forecast interval or budget parameter set based on the divergence degree of the base model or the historical residual distribution.

[0085] Interface module: Encapsulates point predictions and uncertainty sets into standard API outputs for use by EMS or robust optimization scheduling systems.

[0086] The modules communicate with each other through data interfaces, and the base model type or weight calculation method can be flexibly replaced, which facilitates project deployment.

[0087] This embodiment provides a computer-readable storage medium storing a computer program (instructions). When the computer program is executed by a processor (such as a CPU, GPU, FPGA, or embedded controller), it implements the weighted integration method for predicting photovoltaic and wind power generation for multi-energy complementary systems as described in any of the preceding embodiments. The storage medium includes, but is not limited to, ROM, RAM, hard disk, USB flash drive, SD card, optical disk, or cloud storage space.

[0088] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for predicting photovoltaic and wind power generation in multi-energy complementary systems, characterized in that, Includes the following steps: S1. Obtain historical photovoltaic power output data, historical wind power output data and corresponding weather forecast data, and preprocess the historical photovoltaic power output data, the historical wind power output data and the corresponding weather forecast data to obtain a preprocessed historical dataset; S2. Based on the preprocessed historical dataset, identify the weather conditions for photovoltaic power generation and the wind conditions for wind power generation, and generate corresponding condition labels. S3. For the preprocessed historical dataset, use one or more of the following methods: stratified sampling, resampling, or time slicing to generate multiple training subsets with data diversity. S4. On any of the training subsets, train at least two base models with different structures to form a base model set, and obtain the prediction output of each base model. S5. Divide the preprocessed dataset into a training set and a validation set, wherein the training set is used to generate multiple training subsets in step S3. On the validation set, the error index of each base model is calculated according to the working condition category, and the working condition adaptive fusion weights are constructed based on the error index. S6. Based on the current working condition label, call the corresponding fusion weight to perform weighted fusion on the prediction output of each base model to obtain the final point prediction result; S7. Based on the prediction divergence degree and / or historical prediction residual distribution of the base model set, determine the uncertainty scale and construct an uncertainty set for robust optimization, wherein the uncertainty set is a prediction interval set or a budget parameter set. S8. Output the point prediction results and the uncertainty set for robust optimization decision-making of the multi-energy complementary system.

2. The method for predicting photovoltaic and wind power generation for multi-energy complementary systems according to claim 1, characterized in that, The weather conditions for photovoltaic applications include at least three categories: sunny, cloudy, and overcast / rainy, which are classified based on irradiance fluctuations, clear sky index, or cloud cover. The wind conditions for wind power include at least three categories: low wind, medium wind, and strong wind, or two categories: stable and fluctuating, based on wind speed range, wind speed fluctuation rate, or turbulence intensity.

3. The method for predicting photovoltaic and wind power generation for multi-energy complementary systems according to claim 1, characterized in that, In step S3, the specific methods for generating multiple training subsets with data diversity include combinations of at least two of the following: Stratified sampling: Stratified sampling is carried out according to working condition labels, seasons, months, or weekdays / weekends to ensure coverage of each working condition; Resampling: The Bootstrap resampling method is used to construct training subsets with different sample distributions; Time slicing: Divide the original data into multiple training subsets based on time windows to form different time periods.

4. The method for predicting photovoltaic and wind power generation for multi-energy complementary systems according to claim 1, characterized in that, In step S5, the fusion weights for adaptive operation are calculated using an inverse error normalization method: in, For the first On the training subset, the first Individual basic models under working conditions The weights below, For the base model under the working condition The error indicators include at least one of mean absolute error, root mean square error, or mean absolute percentage error. is the regularization coefficient.

5. The method for predicting photovoltaic and wind power generation for multi-energy complementary systems according to claim 1, characterized in that, The base model in step S4 includes, but is not limited to, at least two of the following: generalized regression neural network, extreme learning machine, Elman neural network, long short-term memory network, or extreme gradient boosting model.

6. The method for predicting photovoltaic and wind power generation for multi-energy complementary systems according to claim 1, characterized in that, In step S7: When the set of uncertainties is a set of prediction intervals, it is represented as: in, and These are the lower and upper bounds of the forecast, respectively. When the set of uncertainties is a set of budget parameters, it is represented as: And increase budget constraints: in, Indicates the current moment. Indicates the prediction step size. , To predict the total number of time periods, This represents the actual amount of electricity generated. For point prediction results, To predict bias, The uncertainty scale is determined by the standard deviation of the forecast or the quantile of the historical residual distribution based on the set of base models. For budget parameters, subscript Represents photovoltaic or wind power .

7. The method for predicting photovoltaic and wind power generation for multi-energy complementary systems according to claim 1, characterized in that, In step S1, the meteorological forecast data for photovoltaic power includes at least irradiance characteristics, and the meteorological forecast data for wind power includes at least wind speed characteristics. In step S8, the point prediction results and uncertainty set are output at a time scale of 24-hour and 1-hour resolution.

8. A photovoltaic and wind power generation prediction system for multi-energy complementary systems, used to implement the photovoltaic and wind power generation prediction method for multi-energy complementary systems as described in any one of claims 1 to 7, characterized in that, include: Data processing module: used to perform data acquisition, alignment, cleaning, normalization, and feature construction; Operating condition identification module: used to generate weather operating condition labels for photovoltaic power and wind operating condition labels for wind power. Sub-training set generation module: used to generate multiple training subsets with data diversity; Base Model Training and Inference Module: Used to train multiple base models with different structures and obtain their prediction outputs; Weight learning and fusion module: used to calculate the adaptive weights for working conditions and output the weighted fusion point prediction results; Uncertainty assessment module: used to output forecast intervals or budget set parameters; Interface module: Used to standardize and output the point prediction results and uncertainty set parameters to the robust optimization module or energy management system of the multi-energy complementary system.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the photovoltaic and wind power generation prediction method for multi-energy complementary systems as described in any one of claims 1 to 7.

Images