A photovoltaic power station power prediction method, device, equipment and storage medium

Abstract

The application provides a photovoltaic power station power prediction method, device, equipment and storage medium, the method comprises the following steps: obtaining historical data and processing to obtain a training data set; curve fitting and wavelet decomposition are carried out on the training data set to obtain a feature data set; gradient boosting decision tree and gated recurrent unit network model are trained respectively to obtain a prediction result; according to the number of samples, the prediction result is combined by selecting an integration strategy to obtain the final predicted power. The curve fitting and wavelet decomposition of the application construct the feature data set which fuses physical law and multi-scale time sequence characteristics, and the gradient boosting decision tree and the gated recurrent unit network are trained cooperatively, finally the integration strategy is selected according to the data scale to obtain the power prediction result, realizing the cross-type integration of machine learning, deep learning and physical model, which can effectively improve the prediction accuracy.

Description

Technical Field

[0001] This application relates to the field of photovoltaic power generation prediction technology, and in particular to a method, apparatus, equipment and storage medium for predicting the power of a photovoltaic power plant. Background Technology

[0002] In the field of photovoltaic power generation prediction technology, existing prediction schemes are mainly divided into two categories: physical mechanism-based modeling methods and mathematical statistics methods based on historical data. Among them, physical modeling methods rely heavily on photovoltaic panel parameters and accurate environmental data, resulting in large model errors under complex weather conditions. While mainstream data-driven methods (such as single LSTM models or LightGBM models) can establish the mapping relationship between input and output through data mining, they generally suffer from a technical bottleneck: a single model architecture cannot simultaneously capture the complex nonlinear relationship between weather and power and model the dynamic evolution of time series. Furthermore, their purely data-driven nature makes the model prone to producing prediction outputs that violate the fundamental physical law that "photovoltaic power generation increases monotonically with solar irradiance" when training data is sparse or encounters abnormal weather, thus limiting prediction accuracy and greatly restricting the engineering application value of the prediction results. Summary of the Invention

[0003] This application provides a method, apparatus, equipment, and storage medium for predicting the power of a photovoltaic power plant, which addresses the problem of limited prediction accuracy of a single model in related technologies.

[0004] The first aspect of this application provides a method for predicting the power output of a photovoltaic power plant, the method comprising:

[0005] Historical meteorological data and historical measured power data of photovoltaic power plants are acquired, and the historical meteorological data and historical measured power data are processed to obtain a training dataset;

[0006] The feature dataset is obtained by performing curve fitting and wavelet decomposition on the training dataset;

[0007] A gradient boosting decision tree model is trained based on the feature dataset, and a first prediction result of the gradient boosting decision tree model is generated based on the feature data of the date to be predicted.

[0008] A gated recurrent unit network model is trained based on the feature dataset, and a second prediction result of the gated recurrent unit network model is generated based on the feature data of the date to be predicted.

[0009] Based on the number of samples in the training dataset, a corresponding ensemble strategy is selected to combine the first prediction result and the second prediction result to obtain the predicted power of the photovoltaic power station.

[0010] Optionally, in the first implementation of the first aspect of this application, the step of processing the historical meteorological data and the historical measured power data to obtain the training dataset includes:

[0011] Random sampling and curve fitting are performed on the irradiance data and the historical measured power data in the historical meteorological data to generate a power reference curve;

[0012] Based on the power reference curve, the predicted power value corresponding to each data point in the historical measured power data is determined, and the difference between the measured power value and the predicted power value in the historical measured power data is calculated to obtain the sample residual of each data point.

[0013] Calculate the standard deviation of all the sample residuals, and compare the sample residuals with a positive and negative threshold range set based on the standard deviations to identify the target data points whose sample residuals exceed the positive and negative threshold ranges;

[0014] The target data points are removed from the historical meteorological data and historical measured power data to obtain the training dataset.

[0015] Optionally, in a second implementation of the first aspect of this application, the step of obtaining the feature dataset by performing curve fitting and wavelet decomposition on the training dataset includes:

[0016] The irradiance data in the training dataset is fitted with a polynomial function to the historical measured power data, and a monotonically increasing constraint is applied to the fitting result to generate a monotonically increasing power fitting curve.

[0017] Based on the power fitting curve, the curve prediction value of each sample in the training dataset is determined, and the curve fitting residual of each sample is determined according to the measured power value of the historical measured power data and the curve prediction value.

[0018] Wavelet decomposition is performed on the irradiance time series of historical meteorological data in the training dataset, decomposing the irradiance time series into approximate coefficients reflecting long-term trends and detail coefficients reflecting fluctuations at different time scales.

[0019] Wavelet feature data is obtained by calculating the mean and standard deviation of the approximation coefficients and the energy values ​​and energy ratios of the detail coefficients.

[0020] The basic meteorological features, time features, curve prediction values ​​and curve fitting residuals, and wavelet feature data in the training dataset are integrated to obtain the feature dataset.

[0021] Optionally, in a third implementation of the first aspect of this application, the step of training a gradient boosting decision tree model based on the feature dataset and generating a first prediction result of the gradient boosting decision tree model based on the feature data of the date to be predicted includes:

[0022] The feature dataset is divided into a training subset and a validation subset, and the hyperparameter configuration that minimizes the mean absolute error on the validation subset is determined in a preset hyperparameter space.

[0023] Based on the hyperparameter configuration, an iterative decision tree is constructed on the training subset. In each iteration, the iterative residual between the predicted value and the true value of the training subset is determined. A new decision tree is constructed by fitting the iterative residual until the iteration is completed and a gradient boosting decision tree model is obtained.

[0024] The feature data of the date to be predicted is input into the gradient boosting decision tree model to obtain the first prediction result.

[0025] Optionally, in the fourth implementation of the first aspect of this application, the step of training a gated recurrent unit network model based on the feature dataset and generating a second prediction result of the gated recurrent unit network model based on the feature data of the date to be predicted includes:

[0026] The feature dataset is rearranged in chronological order to construct a time series sample set consisting of samples from consecutive time steps;

[0027] Initialize the sequence processing network structure and train the sequence processing network structure based on the time series sample set; wherein, the sequence processing network structure includes at least two gated recurrent unit layers, at least one random deactivation layer and at least one fully connected layer, as a gated recurrent unit network model to be trained;

[0028] The feature data of the date to be predicted is input into the trained gated recurrent unit network model to obtain the second prediction result.

[0029] Optionally, in the fifth implementation of the first aspect of this application, the step of combining the first prediction result and the second prediction result with the corresponding ensemble strategy selected according to the number of samples in the training dataset to obtain the predicted power of the photovoltaic power station includes:

[0030] The number of samples in the training dataset is compared with a preset first threshold and a second threshold; wherein the first threshold is less than the second threshold.

[0031] When the total number of samples is less than the first quantity threshold, the first prediction result and the second prediction result are weighted and averaged to obtain the predicted power of the photovoltaic power station.

[0032] When the total number of samples is greater than or equal to the first quantity threshold and less than the second quantity threshold, the first prediction result is used as the basic prediction value, and the residual prediction value is obtained through the residual learning of the gated recurrent unit network model. The basic prediction value and the residual prediction value are added together to obtain the predicted power of the photovoltaic power station.

[0033] When the total number of samples is greater than or equal to the second quantity threshold, the first prediction result, the second prediction result, and the third prediction result generated by the power fitting curve are used as meta-features to train the meta-learner, and the predicted power of the photovoltaic power station is output through the meta-learner.

[0034] Optionally, in a sixth implementation of the first aspect of this application, the method further includes:

[0035] The solar altitude angle and solar azimuth angle time series of the date to be predicted are obtained, and combined with the installed capacity and geographical location information of the photovoltaic power station, the theoretical maximum power generation time series of the date to be predicted is calculated.

[0036] The predicted power time series of the photovoltaic power station is compared point by point with the theoretical maximum power generation time series to identify abnormal prediction points where the predicted power of the corresponding predicted power time series exceeds the theoretical maximum power generation at the same time point.

[0037] The predicted power at the abnormal prediction point is replaced with the theoretical maximum power generation to generate the final power prediction sequence.

[0038] A second aspect of this application provides a photovoltaic power plant power prediction device, which is used to implement a photovoltaic power plant power prediction method. The photovoltaic power plant power prediction device includes:

[0039] The acquisition module is used to acquire historical meteorological data and historical measured power data of photovoltaic power plants, and process the historical meteorological data and historical measured power data to obtain a training dataset;

[0040] The feature extraction module is used to obtain the feature dataset by performing curve fitting and wavelet decomposition on the training dataset;

[0041] The first prediction module is used to train a gradient boosting decision tree model based on the feature dataset, and generate the first prediction result of the gradient boosting decision tree model based on the feature data of the date to be predicted.

[0042] The second prediction module is used to train a gated recurrent unit network model based on the feature dataset, and generate a second prediction result of the gated recurrent unit network model based on the feature data of the date to be predicted.

[0043] The processing module is used to select the corresponding integration strategy based on the number of samples in the training dataset to combine the first prediction result and the second prediction result to obtain the predicted power of the photovoltaic power station.

[0044] A third aspect of this application provides an electronic device, including a memory and a processor, wherein the processor is configured to execute a computer program stored in the memory, and when the processor executes the computer program, it implements the steps in the photovoltaic power plant power prediction method provided in the first aspect of this application.

[0045] The fourth aspect of this application provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the steps in the photovoltaic power plant power prediction method provided in the first aspect of this application.

[0046] In summary, the photovoltaic power prediction method, apparatus, equipment, and storage medium provided in this application acquire and process historical data to obtain a training dataset; curve fitting and wavelet decomposition are performed on the training dataset to obtain a feature dataset; gradient boosting decision tree and gated recurrent unit network models are trained respectively to obtain prediction results; and an ensemble strategy is selected based on the sample size to combine the prediction results and obtain the final predicted power. This application constructs a feature dataset integrating physical laws and multi-scale temporal features through curve fitting and wavelet decomposition, and collaboratively trains gradient boosting decision tree and gated recurrent unit network. Finally, an ensemble strategy is adaptively selected based on the data scale to obtain the power prediction result, achieving cross-type integration of machine learning, deep learning, and physical models, which can effectively improve prediction accuracy. Attached Figure Description

[0047] Figure 1 A schematic flowchart illustrating the photovoltaic power prediction method provided in this application embodiment;

[0048] Figure 2 This is a schematic diagram of photovoltaic data cleaning and power reference curve analysis provided in the embodiments of this application;

[0049] Figure 3 This is a schematic diagram of monotonic polynomial curve fitting provided in an embodiment of this application;

[0050] Figure 4 A schematic diagram illustrating the training process of the gradient boosting decision tree model provided in this application embodiment;

[0051] Figure 5 A schematic diagram illustrating the prediction results of the gradient boosting decision tree model provided in this application embodiment;

[0052] Figure 6 A schematic diagram illustrating the training process of the gated recurrent unit network model provided in this application embodiment;

[0053] Figure 7 A schematic diagram of the prediction results of the gated recurrent unit network model provided in the embodiments of this application;

[0054] Figure 8 This is a schematic diagram of the program modules of the photovoltaic power plant power prediction device provided in the embodiments of this application;

[0055] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0056] To make the inventive objectives, features, and advantages of this application more apparent and understandable, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0057] To address the limited prediction accuracy of single models in related technologies, this application provides a method for predicting the power output of photovoltaic power plants, such as... Figure 1 This is a flowchart illustrating the photovoltaic power plant power prediction method provided in this embodiment. The photovoltaic power plant power prediction method includes the following steps:

[0058] Step 110: Obtain historical meteorological data and historical measured power data of the photovoltaic power station, process the historical meteorological data and historical measured power data to obtain the training dataset.

[0059] Specifically, when processing historical meteorological data and historical measured power data required for photovoltaic power plant power prediction, the main approach is to utilize basic meteorological variables such as irradiance, temperature, and wind speed, along with the measured output power at the corresponding time points. First, the raw data from sensors and metering equipment are time-aligned and format-standardized. Then, obviously erroneous negative power, over-installed capacity values, or distorted measurement data are removed by cleaning. Next, abnormal samples are identified and removed by combining physical correlations and statistical laws, thus converting the original observation sequence into training data with a unified structure and reliable quality. Based on this, a historical database reflecting the status of the photovoltaic system is established.

[0060] Step 120: Obtain the feature dataset by performing curve fitting and wavelet decomposition on the training dataset.

[0061] Specifically, when performing curve fitting and wavelet decomposition on the training data, a fitting curve for photovoltaic output is constructed based on the inherent physical relationship between irradiance and measured power. A polynomial form is used and a monotonically increasing constraint is applied to ensure that the fitting result conforms to the physical characteristic that power does not decrease when irradiance increases. The difference between the fitted value and the measured value is calculated to extract the features of the system deviating from the ideal conditions. Subsequently, wavelet decomposition is performed on the irradiance time series, which is decomposed into an approximate component reflecting the gradual trend of solar radiation and a multi-scale detail component reflecting cloud disturbance. Furthermore, energy distribution and statistical descriptive quantities are extracted so that the input information simultaneously covers the overall change and short-term fluctuation features, thereby improving the model's identifiability.

[0062] Step 130: Train the gradient boosting decision tree model based on the feature dataset, and generate the first prediction result of the gradient boosting decision tree model based on the feature data of the date to be predicted.

[0063] Specifically, a gradient boosting decision tree model is trained based on feature data. By associating feature inputs with target power, a tree model structure based on residual iterative learning is adopted, so that the tree generated in each round fits the error of the previous round, thereby gradually improving the overall prediction accuracy. The optimal tree depth, learning rate and number of leaf nodes are determined by a hyperparameter search mechanism, so that the model has high efficiency and robustness in mining complex nonlinear relationships. After training is completed, the features of the date to be predicted are fed into the model to generate the predicted power based on the gradient boosting mechanism.

[0064] Step 140: Train a gated recurrent unit network model based on the feature dataset, and generate a second prediction result of the gated recurrent unit network model based on the feature data of the date to be predicted.

[0065] Specifically, a gated recurrent unit network is trained based on data with the same features. By inputting time series samples into a multi-layer gated recurrent structure in chronological order, the network uses update and reset gates to extract short-term changes, trend changes, and lag effects. During sequence unfolding, it captures dependencies across multiple time steps. Then, by combining random deactivation and fully connected layers, a deep structure capable of handling nonlinear dynamic characteristics is formed, enabling the model to acquire temporal sensitivity to cloud movement speed, irradiance fluctuations, and temperature disturbances. Subsequently, after providing continuous input of the date to be predicted, another set of prediction results based on the time-series learning mechanism is output.

[0066] Step 150: Select the corresponding ensemble strategy based on the number of samples in the training dataset to combine the first prediction result and the second prediction result to obtain the predicted power of the photovoltaic power station.

[0067] Specifically, after obtaining the outputs of the two types of models, an appropriate integration method is selected based on the scale of the training data: when the amount of data is small, a weighted combination is used to avoid overfitting caused by complex structures; when the scale is medium, a sequence network is used to learn gradients to improve the residuals of the model and improve the overall accuracy; and when the amount of data is sufficient, a meta-learner is trained using meta-features containing the prediction results of the two types of predictions and the prediction values ​​of the fitted curves, so that the information of multiple models is integrated in a higher-level manner, thereby generating a more stable final predicted power and forming a complete short-term power prediction process for photovoltaic power plants.

[0068] In one optional implementation of this embodiment, the step of processing historical meteorological data and historical measured power data to obtain a training dataset includes: randomly sampling and curve fitting the irradiance data and historical measured power data in the historical meteorological data to generate a power baseline curve; determining the predicted power value corresponding to each data point in the historical measured power data based on the power baseline curve, and subtracting the measured power value from the predicted power value to obtain the sample residual for each data point; calculating the standard deviation of all sample residuals, and comparing the sample residuals with a positive and negative threshold range set based on the standard deviation to identify the target data points corresponding to the sample residuals exceeding the positive and negative threshold range; and removing the target data points from the historical meteorological data and historical measured power data to obtain the training dataset.

[0069] In this embodiment, as Figure 2The diagram illustrates photovoltaic data cleaning and power baseline curve analysis. The RANSAC (Random Sample Consensus) algorithm is applied to fit historical irradiance and measured power data. RANSAC repeatedly extracts subsamples from observation points and estimates model parameters based on these subsamples to identify and eliminate outliers that do not conform to the main data support, thus obtaining a robust irradiance-power baseline curve. After the power baseline curve is determined, the irradiance of all historical observation points is input into the curve to obtain the predicted power value for each data point. The difference between the predicted power value and the historical measured power value is calculated, forming the sample residual. The sample residual represents the deviation between the actual power output at a given moment and the theoretical power output under ideal lighting conditions. If an observation point is affected by shading, inverter limitations, sensor drift, or partial component failure, the residual will increase significantly. For example, when the irradiance is under strong light conditions but the actual output is lower than the theoretical value due to power plant curtailment, the residual will show a significant negative deviation. Conversely, if sensor errors cause abnormally high power, a large positive deviation will occur. Because the baseline curve is subject to monotonic constraints, the residual distribution directly reflects anomalies, making the differences between outlier and normal points more prominent. After obtaining all residuals, to determine which observation points should be removed from the training set, the standard deviation of all residuals needs to be calculated. The standard deviation characterizes the dispersion of the residuals within the overall distribution. When residual fluctuations are close to the normal physical response range, their values ​​are smaller, while outlier points, due to excessive relative deviation, significantly increase their position in the distribution. By multiplying the standard deviation by a fixed coefficient to establish positive and negative threshold intervals, for example, using 1.5 times the standard deviation in both directions as boundaries, a range that is relatively tolerant of normal fluctuations but capable of identifying outliers can be formed. If the residual of a point exceeds this range, it indicates that the point is inconsistent with the overall pattern and belongs to the target data point. After the target data point is identified, it is simultaneously removed from historical meteorological and historical power data to generate a training dataset that does not contain outlier samples. The removed data points will no longer participate in subsequent model training, allowing the modeling process to be based on more realistic physical responses and more stable power output patterns, thereby improving the reliability and generalization ability of subsequent feature engineering, gradient boosting model training, and sequence neural network learning.

[0070] In one optional implementation of this embodiment, the step of obtaining a feature dataset by performing curve fitting and wavelet decomposition on the training dataset includes: performing polynomial function fitting on the irradiance data and historical measured power data in the training dataset, and applying a monotonically increasing constraint to the fitting result to generate a monotonically increasing power fitting curve; determining the curve prediction value of each sample in the training dataset based on the power fitting curve, and determining the curve fitting residual of each sample according to the measured power value and the curve prediction value of the historical measured power data; performing wavelet decomposition on the irradiance time series of historical meteorological data in the training dataset, decomposing the irradiance time series into approximation coefficients reflecting long-term trends and detail coefficients reflecting fluctuations at different time scales; obtaining wavelet feature data by calculating the mean and standard deviation of the approximation coefficients and the energy value and energy ratio of the detail coefficients; and integrating the basic meteorological features, time features, curve prediction values ​​and curve fitting residuals, and the wavelet feature data in the training dataset to obtain the feature dataset.

[0071] In this embodiment, a mathematical mapping between irradiance and photovoltaic module output power is first established, allowing the nonlinear response relationship between the two to be expressed parametrically. A polynomial function, with irradiance as the independent variable, constructs a quadratic or cubic expression using a set of coefficients to be determined. This function curve can characterize the power increase caused by enhanced illumination and the slow rise in the low irradiance range, effectively approximating the actual operating curve of the photovoltaic module. A monotonically increasing constraint is introduced during the solution process to prevent the derivative of the fitted function from becoming negative, thus maintaining the physical property that "power does not decrease with increasing irradiance." The monotonic constraint can be achieved through order-preserving regression, a type of optimization method that forces the output sequence to satisfy monotonicity when solving the regression function. By adjusting the fitted values ​​or slopes of each interval, local reverse fluctuations caused by random disturbances are reduced, resulting in a continuously rising curve. Figure 3The diagram illustrates a monotonic polynomial curve fitting process. After the power fitting curve is determined, the irradiance value of each training sample is input into the curve to obtain the predicted curve value. Then, the curve fitting residual is defined based on the difference between the predicted curve value and the historical measured power value, giving each data point a clear characterization of power deviation. If an observation point is temporarily obscured by cloud shadows, its measured power will be significantly lower than the predicted curve value, resulting in a large negative offset in the residual. If sensor drift causes artificially high power at a point, the residual will show a large positive offset. After completing the polynomial fitting, wavelet decomposition is performed on the irradiance time series, breaking down the irradiance variation process into a multi-level time-frequency structure. Wavelet decomposition is a mathematical method that integrates time positioning and frequency decomposition capabilities. By convolving the original time series with a set of wavelet functions with scaling and translation capabilities, long-period smooth changes can be separated into approximate coefficients, and rapid changes, jumps, and local perturbations can be separated into multiple detail coefficients. The approximation coefficients express the overall trend of irradiance evolution over time, such as the gradual increase from sunrise to noon and the gradual decrease in the afternoon. Detail coefficients, on the other hand, capture rapid processes at different scales, such as the passage of short-term cloud clusters, intensity oscillations, and sharp rises and falls. For example, when large-scale clouds pass by, strong fluctuations are observed in the lower-frequency detail coefficients, while the higher-frequency detail coefficients capture slight cloud shadow disturbances. The mean and standard deviation of the approximation coefficients describe the overall illumination level and stability, while calculating the energy value and energy ratio of the detail coefficients quantifies the contribution of disturbances at different frequency bands to irradiance changes, enabling a digital representation of the multi-scale meteorological structure. After the curve predictions, curve fitting residuals, and wavelet features generated by polynomial fitting are formed, they are integrated with the basic meteorological and temporal features in the training dataset. This ensures that each training sample includes information such as irradiance, temperature, wind speed, time location, ideal response power, deviation degree, and multi-scale structure from wavelet analysis, allowing the prediction model to simultaneously utilize physical constraints, temporal patterns, and illumination disturbance features.

[0072] In one optional implementation of this embodiment, the steps of training a gradient boosting decision tree model based on a feature dataset and generating a first prediction result of the gradient boosting decision tree model based on the feature data of the date to be predicted include: dividing the feature dataset into a training subset and a validation subset, and determining the hyperparameter configuration that minimizes the mean absolute error on the validation subset in a preset hyperparameter space; constructing an iterative decision tree on the training subset based on the hyperparameter configuration, and determining the iterative residual between the predicted value and the true value of the training subset in each iteration, constructing a new decision tree by fitting the iterative residual, until the iteration is completed and a gradient boosting decision tree model is obtained; and inputting the feature data of the date to be predicted into the gradient boosting decision tree model to obtain the first prediction result.

[0073] In this embodiment, when modeling the feature dataset, it is divided into a training subset and a validation subset. By dividing the overall data into two parts in a chronological or random manner, the model learns the mapping relationship on the training subset and evaluates its performance on the validation subset, thereby determining whether the model parameters are stable. Since photovoltaic power prediction has strong temporal correlation, maintaining temporal continuity during the partitioning process can reduce future information leakage and make the model closer to the real business environment. After partitioning, the optimal parameter combination needs to be searched in a preset hyperparameter space. The hyperparameter space refers to a set of external control variables of the model, including the learning rate, the maximum depth of the tree, the number of leaf nodes in each tree, and the subsampling ratio used to control model complexity. By traversing multiple combinations in the hyperparameter space and using the mean absolute error as an evaluation metric, the prediction bias of the model on the validation subset can be quantified, allowing the optimal hyperparameter configuration to be determined with the goal of minimizing error. For example, if the learning rate is set too high, the model will become overly sensitive to single-step residual signals during training, leading to oscillations. Conversely, if the learning rate is too low, convergence may be too slow, significantly increasing the mean absolute error on the validation set. Therefore, feedback from the validation set can accurately select a more suitable learning rate range. After obtaining the optimal hyperparameter configuration, the training subset is input into the gradient boosting framework to begin building a model composed of multiple decision trees. Gradient boosting is an ensemble learning method based on "gradually approximating the true target value." It improves overall performance by sequentially combining multiple weak learners, typically using a decision tree structure. During the construction process, an initial model first generates a set of preliminary predictions. This initial model can be simply taken as the average of the target values ​​in the training set, providing a clear starting point for subsequent updates. Subsequently, the first iteration residual is calculated based on the difference between the true value and the initial prediction. The iteration residual refers to the offset between the true target and the current model prediction. The core idea of ​​gradient boosting is to gradually approximate the true state by continuously fitting this offset. The iterative residuals are input into the weak learner as the new training target, constructing a decision tree that reflects the residual structure. The learning rate controls the influence of this tree on the overall model, thus avoiding performance instability caused by large updates. If the residuals exhibit different characteristic intervals in a certain iteration—for example, smaller residuals in sunny, high-irradiance intervals and larger residuals in intervals with frequent cloud disturbances—the newly constructed tree will generate more splitting nodes in the cloud disturbance intervals to enhance the fitting ability, making the model better able to capture the impact of irradiance fluctuations. Subsequently, the predictions generated by the new tree update the overall model, making the overall predictions further approach the true values. The above iterative process continues until a set number of iterations is reached or the validation set error stops decreasing, at which point a complete gradient boosting decision tree model is obtained. Figure 4The diagram illustrates the training process of a gradient boosting decision tree model. After training, the feature data for the date to be predicted is input into the model, allowing it to infer the future irradiance-power relationship using the previously learned structure. Because the model has learned comprehensive patterns based on fundamental meteorological features, multi-scale fluctuation features, and power deviation features during training, it can automatically select appropriate decision paths based on the input features during prediction. For example, when the input irradiance features exhibit significant high-frequency fluctuations, the model will use the tree branch corresponding to the high-frequency fluctuations formed during training, thus obtaining prediction results that are more sensitive to rapid cloud changes; while when the input features show a steady upward or downward trend, the model will move towards the branch representing clear sky trends, making the prediction results more consistent with the photovoltaic behavior during that period. Through this mechanism, the gradient boosting model can flexibly adjust the prediction path according to the input conditions and ultimately output as shown in the diagram. Figure 5 The first prediction result shown makes the photovoltaic power plant power prediction more adaptable and accurate, forming a consistent logical structure in the model inference process and the training phase, thereby ensuring the prediction quality and reliability.

[0074] In one optional implementation of this embodiment, the step of training a gated recurrent unit network model based on a feature dataset and generating a second prediction result of the gated recurrent unit network model based on the feature data of the date to be predicted includes: rearranging the feature dataset in chronological order to construct a time series sample set composed of samples from consecutive time steps; initializing a sequence processing network structure and training the sequence processing network structure based on the time series sample set; wherein the sequence processing network structure includes at least two gated recurrent unit layers, at least one random deactivation layer, and at least one fully connected layer as the gated recurrent unit network model to be trained; and inputting the feature data of the date to be predicted into the trained gated recurrent unit network model to obtain the second prediction result.

[0075] In this embodiment, when constructing a sequence network model based on gated recurrent units using the feature dataset, all feature samples are first rearranged in chronological order. This transforms the feature structure, originally stored as a single point, into a continuous sequence that reflects sequential relationships, allowing the trend of photovoltaic power evolution over time, short-term disturbances, and hysteresis effects to be expressed at the data input. During sequence construction, several adjacent time points are combined into continuous windows, making each window a complete time series sample. This sample contains multi-dimensional variables such as basic meteorological features, curve prediction values, curve fitting residuals, and wavelet features, enabling the model to simultaneously capture physical trends, deviations, and multi-scale disturbance structures when the sequence is unfolded. For example, if a sequence sample is constructed using the past 24 time points, the sequence can represent the complete cyclic behavior of irradiance and power within a day, thus providing a sample structure with sufficient contextual information for the gated recurrent units. After completing the sequence construction, the sequence processing network structure is initialized, and a deep network containing multiple layers of gated recurrent units, random deactivation layers, and fully connected layers is established. Gated recurrent units (ROUs) are a type of recurrent neural structure used to process time series data. The update gate and reset gate are two core components. The update gate controls the extent to which the current unit state inherits the hidden state from the previous time step, while the reset gate controls how the current input is integrated with historical states, enabling the network to capture long-term dependencies and short-term changes. By stacking at least two layers of GUs, the network can identify local rapid fluctuations at lower layers and overall trends and slow perturbations at higher layers. This allows the model to distinguish the differences in response of photovoltaic systems affected by high-frequency cloud shadows versus low-frequency weather changes. Random deactivation layers are structures that randomly shield some neurons during training. By randomly zeroing out some activation values ​​during forward propagation, the network avoids over-reliance on specific dimensional features, thereby reducing sensitivity and enhancing the model's generalization ability. For example, in a training batch, after some wavelet energy features are randomly deactivated, the model still needs to learn temporal relationships from the remaining features. This reinforcement learning approach makes the model more stable when facing feature perturbations during deployment. The fully connected layer, located at the end of the network, integrates the temporal information output by the gated recurrent units into a single-step prediction. Through linear combination and nonlinear activation, it maps the temporal features learned by the multi-layer hidden structure into power prediction values. For example... Figure 6The diagram illustrates the training process of a gated recurrent unit (GRU) network model. After initializing the sequence structure and network architecture, the entire model is trained using a time-series sample set. The network parameters are optimized through continuous forward and backward propagation, gradually reducing the loss function. During training, the GRU processes feature inputs step-by-step, combining the hidden state from the previous time step with the features of the current time step. The update gate adjusts the historical memory weights, and the reset gate controls the proportion of new information injection, ensuring that the internal state at each time step comprehensively expresses both recent fluctuations and long-term trends. For example, during a period of continuous cloud disturbance, the reset gate enhances the impact of the current input, while during a period of clear and stable weather, the update gate strengthens the inheritance of long-term trends. This allows the model to automatically adapt to the temporal representation of different weather processes, ultimately forming a complete parameter set. As training progresses, the network gradually learns the dynamic correlation between irradiance changes, wavelet energy distribution, power deviation structure, and time factors, enabling it to extract patterns from historical data and infer future conditions. Once the training reaches convergence, the feature data for the date to be predicted is input into the network in the same format as during training. This allows the network to propagate the input information gradually over time, and the final fully connected layer outputs the data as shown in the diagram. Figure 7 The second prediction result is shown.

[0076] In an optional implementation of this embodiment, the step of combining the first prediction result and the second prediction result with the corresponding ensemble strategy selected based on the number of samples in the training dataset to obtain the predicted power of the photovoltaic power station includes: comparing the number of samples in the training dataset with a preset first quantity threshold and a second quantity threshold; wherein the first quantity threshold is less than the second quantity threshold; when the total number of samples is less than the first quantity threshold, performing a weighted average of the first prediction result and the second prediction result to obtain the predicted power of the photovoltaic power station; when the total number of samples is greater than or equal to the first quantity threshold and less than the second quantity threshold, using the first prediction result as the basic prediction value, and obtaining the residual prediction value through residual learning of the gated recurrent unit network model, adding the basic prediction value and the residual prediction value to obtain the predicted power of the photovoltaic power station; when the total number of samples is greater than or equal to the second quantity threshold, using the first prediction result, the second prediction result, and the third prediction result generated by the power fitting curve as meta-features to train a meta-learner, and outputting the predicted power of the photovoltaic power station through the meta-learner.

[0077] Specifically, in photovoltaic power plant power prediction, the choice of prediction method is closely related to the sample size of the training dataset. By comparing the sample size of the training dataset with a preset first and second threshold, an adaptive prediction strategy can be achieved. The optimal ensemble strategy is dynamically selected based on the sample size of the training dataset, thus achieving an optimal balance between prediction accuracy and model stability in three different scenarios: scarce, moderate, and abundant data. The first threshold is less than the second threshold. When the total number of samples is lower than the first threshold, it indicates that the historical information available for model training is limited. In this case, using a complex ensemble model can easily lead to overfitting, i.e., the model excessively memorizes noise from the training data and loses its generalization ability. Therefore, in this data-scarce scenario, a computationally efficient and structurally simple weighted average strategy is adopted. Specifically, on a validation set independent of the training process, the optimal weight coefficients for the first prediction result of the gradient boosting decision tree model and the second prediction result of the gated recurrent unit network model need to be determined through grid search or optimization algorithms, with the constraint that the sum of the weights is 1. The final predicted power is the linear superposition of the weighted first and second prediction results. The advantage of this strategy lies in fusing the predictive tendencies of two heterogeneous models with minimal model complexity, effectively avoiding model performance instability caused by insufficient data. When the total number of samples reaches or exceeds the first threshold but remains below the second threshold, it means that the amount of data is sufficient to support more refined model collaboration, at which point the residual learning ensemble strategy is activated. This strategy establishes the first prediction result of the gradient boosting decision tree model as the base prediction value, which is considered to capture the main mapping relationship between power and static features such as weather and time. Subsequently, the gated recurrent unit network model is given a new learning objective: instead of directly predicting power, it aims to fit the residual between the base prediction value and the true power value. This residual represents the complex temporal fluctuations and nonlinear patterns that the gradient boosting decision tree model fails to explain. The gated recurrent unit network, with its internal gating mechanism (update gate and reset gate), can effectively capture long-term dependencies in the time series, thereby accurately learning these residual patterns. The final predicted power consists of the base prediction value plus the residual prediction value output by the gated recurrent unit network, forming a cascaded and complementary prediction process. When the sample size is further expanded, reaching or exceeding the second quantity threshold, it indicates the availability of massive and diverse historical data, sufficient to train a more complex but potentially more powerful stacked ensemble model. In this large-sample scenario, the ensemble framework is upgraded to a two-layer structure. In the first layer, the first prediction result generated by the gradient boosting decision tree model, the second prediction result generated by the gated recurrent unit network model, and the third prediction result directly calculated from the power fitting curve (a baseline curve generated by polynomial fitting and order-preserving regression that strictly satisfies the physical law of monotonically increasing power with irradiance) are collectively used as meta-features.These meta-features combine different perspectives and predictive intelligence from machine learning models, deep learning models, and physical constraint models. The second layer introduces a meta-learner, typically a relatively simple shallow neural network, whose input is the meta-feature vector composed of the three prediction results mentioned above. This meta-learner learns on the training set how to optimally weight and combine these meta-features, and its output is the final, more accurate predicted power of the photovoltaic power plant. This stacked integration strategy automatically discovers the complex interactions and dependencies between the prediction results of different base models through the meta-learner, thereby maximizing prediction performance under conditions of sufficient data.

[0078] In one optional implementation of this embodiment, the solar altitude angle and solar azimuth angle time series for the date to be predicted are obtained, and combined with the installed capacity and geographical location information of the photovoltaic power station, the theoretical maximum power generation time series for the date to be predicted is calculated; the predicted power time series of the photovoltaic power station is compared point by point with the theoretical maximum power generation time series to identify abnormal prediction points where the predicted power of the corresponding predicted power time series exceeds the theoretical maximum power generation at the same time point; the predicted power of the abnormal prediction points is replaced with the theoretical maximum power generation to generate the final power prediction series.

[0079] In this embodiment, the solar altitude angle is defined as the angle between the sun's rays and the ground plane, ranging from -90 degrees to +90 degrees, reaching its maximum value at noon; the solar azimuth angle represents the angle between the sun's projection onto the ground plane and true north, ranging from 0 degrees to 360 degrees. Based on the specific latitude and longitude coordinates of the photovoltaic power station and the year and time series information of the date to be predicted, an astronomical equation including parameters such as declination angle, hour angle, and solar time is iteratively solved to finally generate a solar altitude angle and solar azimuth angle time series that perfectly match the predicted time resolution. After obtaining accurate celestial trajectory data, combined with the installed capacity of the photovoltaic power station—that is, the rated maximum power generation value of the power station—the theoretical maximum power generation time series is calculated through a photovoltaic power generation physical model. This calculation process needs to comprehensively consider factors such as the geometric relationship of the sun's position, the tilt angle and azimuth angle configuration of the photovoltaic panels, and atmospheric optical quality. In particular, it is necessary to introduce an incident angle correction coefficient to quantify the impact of the angle between the sun's rays and the normal to the photovoltaic panel on energy reception efficiency. The photovoltaic panel achieves the highest reception efficiency when sunlight strikes it perpendicularly. As the incident angle increases, cosine energy loss occurs. The theoretical maximum power generation can be calculated as follows: theoretical power equals the product of installed capacity and normal direct irradiance, multiplied by a composite trigonometric function formed by the sine of the solar altitude angle and the cosine of the difference in azimuth angle. Finally, a temperature correction coefficient and a system efficiency coefficient are introduced for calibration. Through this series of calculations, a theoretical power upper limit curve conforming to the physical characteristics of the power station and local astronomical patterns can be obtained. After establishing the theoretical benchmark, the original predicted power time series output by the data-driven model is compared with the theoretical maximum power generation time series at each time point. Abnormal data points where the predicted power value exceeds the theoretical maximum power generation value at the same time point are accurately identified. These anomalies often arise because the machine learning model learns data patterns containing measurement errors or abnormal operating states during training, leading to over-predictions that violate physical laws under specific meteorological conditions. For identified abnormal prediction points, the system activates an automatic correction mechanism, replacing the predicted power value with the corresponding theoretical maximum power generation value. This replacement operation maintains the integrity of prediction results for non-anomalies, intervening only for data points that violate physical laws. The final power prediction sequence retains the intelligent prediction characteristics of the data-driven model while strictly adhering to physical constraints. This correction mechanism effectively solves the power over-prediction problem that may arise from purely data-driven methods. For example, during rapid changes from cloudy to sunny weather, the model may generate power predictions exceeding the physical limits of the power plant due to learning from historical anomalies. This physical constraint correction ensures that the prediction results always remain within reasonable physical boundaries.

[0080] According to the photovoltaic power prediction method provided in this application, historical data is acquired and processed to obtain a training dataset; curve fitting and wavelet decomposition are performed on the training dataset to obtain a feature dataset; gradient boosting decision tree and gated recurrent unit network models are trained respectively to obtain prediction results; and an ensemble strategy is selected based on the number of samples to combine the prediction results to obtain the final predicted power. This application constructs a feature dataset that integrates physical laws and multi-scale temporal features through curve fitting and wavelet decomposition, and collaboratively trains gradient boosting decision tree and gated recurrent unit network. Finally, the ensemble strategy is adaptively selected based on the data scale to obtain the power prediction result, realizing cross-type integration of machine learning, deep learning, and physical models, which can effectively improve prediction accuracy.

[0081] Figure 2 This application provides a photovoltaic power plant power prediction device, which can be used to implement the photovoltaic power plant power prediction method in the aforementioned embodiments. Figure 2 As shown, the photovoltaic power station power prediction device mainly includes:

[0082] The acquisition module 10 is used to acquire historical meteorological data and historical measured power data of photovoltaic power plants, and to process the historical meteorological data and historical measured power data to obtain a training dataset.

[0083] The feature extraction module 20 is used to obtain the feature dataset by performing curve fitting and wavelet decomposition on the training dataset;

[0084] The first prediction module 30 is used to train a gradient boosting decision tree model based on the feature dataset and generate the first prediction result of the gradient boosting decision tree model based on the feature data of the date to be predicted.

[0085] The second prediction module 40 is used to train a gated recurrent unit network model based on the feature dataset and generate a second prediction result of the gated recurrent unit network model based on the feature data of the date to be predicted.

[0086] The processing module 50 is used to select the corresponding integration strategy based on the number of samples in the training dataset to combine the first prediction result and the second prediction result to obtain the predicted power of the photovoltaic power station.

[0087] In one optional implementation of this embodiment, the acquisition module is specifically used for: randomly sampling and curve fitting the irradiance data and historical measured power data in historical meteorological data to generate a power reference curve; determining the predicted power value corresponding to each data point in the historical measured power data based on the power reference curve, and subtracting the measured power value and the predicted power value from the historical measured power data to obtain the sample residual of each data point; calculating the standard deviation of all sample residuals, and comparing the sample residuals with a positive and negative threshold range set based on the standard deviation to identify the target data points corresponding to the sample residuals exceeding the positive and negative threshold range; and removing the target data points from the historical meteorological data and historical measured power data to obtain a training dataset.

[0088] In one optional implementation of this embodiment, the feature extraction module is specifically used for: performing polynomial function fitting on the irradiance data and historical measured power data in the training dataset, and applying a monotonically increasing constraint to the fitting result to generate a monotonically increasing power fitting curve; determining the curve prediction value of each sample in the training dataset based on the power fitting curve, and determining the curve fitting residual of each sample based on the measured power value and the curve prediction value of the historical measured power data; performing wavelet decomposition on the irradiance time series of historical meteorological data in the training dataset, decomposing the irradiance time series into approximation coefficients reflecting long-term trends and detail coefficients reflecting fluctuations at different time scales; obtaining wavelet feature data by calculating the mean and standard deviation of the approximation coefficients and the energy value and energy ratio of the detail coefficients; and integrating the basic meteorological features, time features, curve prediction values ​​and curve fitting residuals, and the wavelet feature data in the training dataset to obtain a feature dataset.

[0089] In one optional implementation of this embodiment, the first prediction module is specifically used to: divide the feature dataset into a training subset and a validation subset, and determine the hyperparameter configuration that minimizes the mean absolute error on the validation subset in a preset hyperparameter space; construct an iterative decision tree on the training subset according to the hyperparameter configuration, and determine the iterative residual between the predicted value and the true value of the training subset in each iteration, and construct a new decision tree by fitting the iterative residual until the iteration is completed to obtain a gradient boosting decision tree model; input the feature data of the date to be predicted into the gradient boosting decision tree model to obtain the first prediction result.

[0090] In one optional implementation of this embodiment, the second prediction module is specifically used for: rearranging the feature dataset in chronological order to construct a time series sample set composed of samples from consecutive time steps; initializing the sequence processing network structure and training the sequence processing network structure based on the time series sample set; wherein the sequence processing network structure includes at least two gated recurrent unit layers, at least one random deactivation layer, and at least one fully connected layer as a gated recurrent unit network model to be trained; and inputting the feature data of the date to be predicted into the trained gated recurrent unit network model to obtain the second prediction result.

[0091] In an optional implementation of this embodiment, the processing module is specifically used to: compare the number of samples in the training dataset with a preset first quantity threshold and a second quantity threshold; wherein the first quantity threshold is less than the second quantity threshold; when the total number of samples is less than the first quantity threshold, perform a weighted average of the first prediction result and the second prediction result to obtain the predicted power of the photovoltaic power station; when the total number of samples is greater than or equal to the first quantity threshold and less than the second quantity threshold, use the first prediction result as the basic prediction value, and obtain the residual prediction value through residual learning of the gated recurrent unit network model, and add the basic prediction value and the residual prediction value to obtain the predicted power of the photovoltaic power station; when the total number of samples is greater than or equal to the second quantity threshold, use the first prediction result, the second prediction result, and the third prediction result generated by the power fitting curve as meta-features to train the meta-learner, and output the predicted power of the photovoltaic power station through the meta-learner.

[0092] In an optional embodiment of this example, the processing module is further configured to: obtain the solar altitude angle and solar azimuth angle time series for the date to be predicted, and calculate the theoretical maximum power generation time series for the date to be predicted by combining the installed capacity and geographical location information of the photovoltaic power station; compare the predicted power time series of the photovoltaic power station with the theoretical maximum power generation time series point by point, identify abnormal prediction points where the predicted power of the corresponding predicted power time series exceeds the theoretical maximum power generation at the same time point; replace the predicted power of the abnormal prediction points with the theoretical maximum power generation, and generate the final power prediction series.

[0093] According to the photovoltaic power prediction device provided in this application, historical data is acquired and processed to obtain a training dataset; curve fitting and wavelet decomposition are performed on the training dataset to obtain a feature dataset; gradient boosting decision tree and gated recurrent unit network models are trained respectively to obtain prediction results; and an ensemble strategy is selected based on the number of samples to combine the prediction results to obtain the final predicted power. This application constructs a feature dataset that integrates physical laws and multi-scale temporal features through curve fitting and wavelet decomposition, and collaboratively trains gradient boosting decision tree and gated recurrent unit network. Finally, an ensemble strategy is adaptively selected based on the data scale to obtain the power prediction result, realizing cross-type integration of machine learning, deep learning, and physical models, which can effectively improve prediction accuracy.

[0094] According to the scheme provided in this application Figure 9 An electronic device is provided as an embodiment of this application. This electronic device can be used to implement the photovoltaic power plant power prediction method in the foregoing embodiments, and mainly includes:

[0095] The system includes a memory 901, a processor 902, and a computer program 903 stored in the memory 901 and executable on the processor 902. The memory 901 and the processor 902 are connected via communication. When the processor 902 executes the computer program 903, it implements the photovoltaic power plant power prediction method described in the foregoing embodiments. The number of processors can be one or more.

[0096] The memory 901 can be a high-speed random access memory (RAM) or a non-volatile memory, such as a disk storage device. The memory 901 is used to store executable program code, and the processor 902 is coupled to the memory 901.

[0097] Furthermore, embodiments of this application also provide a computer-readable storage medium, which may be disposed in the electronic device described in the above embodiments, and the computer-readable storage medium may be as described above. Figure 3 The memory in the illustrated embodiment.

[0098] The computer-readable storage medium stores a computer program that, when executed by a processor, implements the photovoltaic power plant power prediction method described in the foregoing embodiments. Furthermore, the computer-readable storage medium can also be any medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), RAM, magnetic disk, or optical disk.

[0099] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0100] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0101] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for predicting the power output of a photovoltaic power plant, characterized in that, include: Historical meteorological data and historical measured power data of photovoltaic power plants are acquired, and the historical meteorological data and historical measured power data are processed to obtain a training dataset; The feature dataset is obtained by performing curve fitting and wavelet decomposition on the training dataset; A gradient boosting decision tree model is trained based on the feature dataset, and a first prediction result of the gradient boosting decision tree model is generated based on the feature data of the date to be predicted; that is, the feature dataset is divided into a training subset and a validation subset, and the hyperparameter configuration with the minimum mean absolute error on the validation subset is determined in a preset hyperparameter space. According to the hyperparameter configuration, an iterative decision tree is constructed on the training subset. In each iteration, the iterative residual between the predicted value and the true value of the training subset is determined. A new decision tree is constructed by fitting the iterative residual until the iteration is completed to obtain the gradient boosting decision tree model. The feature data of the date to be predicted is input into the gradient boosting decision tree model to obtain the first prediction result. A gated recurrent unit network (GRN) model is trained based on the feature dataset, and a second prediction result of the GRN model is generated based on the feature data of the date to be predicted. Specifically, the feature dataset is rearranged chronologically to construct a time series sample set composed of samples from consecutive time steps. A sequence processing network structure is initialized and trained based on the time series sample set. The sequence processing network structure includes at least two gated recurrent unit layers, at least one randomly deactivated layer, and at least one fully connected layer, serving as the GRN model to be trained. The feature data of the date to be predicted is input into the trained GRN model to obtain the second prediction result. Based on the sample size of the training dataset, a corresponding ensemble strategy is selected to combine the first prediction result and the second prediction result to obtain the predicted power of the photovoltaic power station. Specifically, the sample size of the training dataset is compared with a preset first and second quantity thresholds, wherein the first quantity threshold is less than the second quantity threshold. When the total number of samples is less than the first quantity threshold, a weighted average is taken of the first and second prediction results to obtain the predicted power of the photovoltaic power station. When the total number of samples is greater than or equal to the first quantity threshold and less than the second quantity threshold, the first prediction result is used as the base prediction value, and a residual prediction value is obtained through residual learning of the gated recurrent unit network model. The base prediction value and the residual prediction value are added to obtain the predicted power of the photovoltaic power station. When the total number of samples is greater than or equal to the second quantity threshold, the first prediction result, the second prediction result, and the third prediction result generated from the power fitting curve are used as meta-features to train a meta-learner, and the predicted power of the photovoltaic power station is output through the meta-learner.

2. The photovoltaic power plant power prediction method according to claim 1, characterized in that, The step of processing the historical meteorological data and the historical measured power data to obtain the training dataset includes: Random sampling and curve fitting are performed on the irradiance data and the historical measured power data in the historical meteorological data to generate a power reference curve; Based on the power reference curve, the predicted power value corresponding to each data point in the historical measured power data is determined, and the difference between the measured power value and the predicted power value in the historical measured power data is calculated to obtain the sample residual of each data point. Calculate the standard deviation of all the sample residuals, and compare the sample residuals with a positive and negative threshold range set based on the standard deviations to identify the target data points whose sample residuals exceed the positive and negative threshold ranges; The target data points are removed from the historical meteorological data and historical measured power data to obtain the training dataset.

3. The photovoltaic power plant power prediction method according to claim 1, characterized in that, The step of obtaining the feature dataset by performing curve fitting and wavelet decomposition on the training dataset includes: The irradiance data in the training dataset is fitted with a polynomial function to the historical measured power data, and a monotonically increasing constraint is applied to the fitting result to generate a monotonically increasing power fitting curve. Based on the power fitting curve, the curve prediction value of each sample in the training dataset is determined, and the curve fitting residual of each sample is determined according to the measured power value of the historical measured power data and the curve prediction value. Wavelet decomposition is performed on the irradiance time series of historical meteorological data in the training dataset, decomposing the irradiance time series into approximate coefficients reflecting long-term trends and detail coefficients reflecting fluctuations at different time scales. Wavelet feature data is obtained by calculating the mean and standard deviation of the approximation coefficients and the energy values ​​and energy ratios of the detail coefficients. The basic meteorological features, time features, curve prediction values ​​and curve fitting residuals, and wavelet feature data in the training dataset are integrated to obtain the feature dataset.

4. The photovoltaic power plant power prediction method according to claim 1, characterized in that, The method further includes: The solar altitude angle and solar azimuth angle time series of the date to be predicted are obtained, and combined with the installed capacity and geographical location information of the photovoltaic power station, the theoretical maximum power generation time series of the date to be predicted is calculated. The predicted power time series of the photovoltaic power station is compared point by point with the theoretical maximum power generation time series to identify abnormal prediction points where the predicted power of the corresponding predicted power time series exceeds the theoretical maximum power generation at the same time point. The predicted power at the abnormal prediction point is replaced with the theoretical maximum power generation to generate the final power prediction sequence.

5. A photovoltaic power plant power prediction device, characterized in that, The photovoltaic power plant power prediction device is used to implement the photovoltaic power plant power prediction method according to claim 1, and the photovoltaic power plant power prediction device includes: The acquisition module is used to acquire historical meteorological data and historical measured power data of photovoltaic power plants, and process the historical meteorological data and historical measured power data to obtain a training dataset; The feature extraction module is used to obtain the feature dataset by performing curve fitting and wavelet decomposition on the training dataset; The first prediction module is used to train a gradient boosting decision tree model based on the feature dataset, and generate the first prediction result of the gradient boosting decision tree model based on the feature data of the date to be predicted. The second prediction module is used to train a gated recurrent unit network model based on the feature dataset, and generate a second prediction result of the gated recurrent unit network model based on the feature data of the date to be predicted. The processing module is used to select the corresponding integration strategy based on the number of samples in the training dataset to combine the first prediction result and the second prediction result to obtain the predicted power of the photovoltaic power station.

6. An electronic device, characterized in that, Includes memory and processor, of which: The processor is used to execute computer programs stored in the memory; When the processor executes the computer program, it implements the steps in the photovoltaic power plant power prediction method according to any one of claims 1 to 4.

7. 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 steps in the photovoltaic power plant power prediction method according to any one of claims 1 to 4.

Images