Photovoltaic power station power prediction method, device, equipment and storage medium

Abstract

The invention provides a photovoltaic power station power prediction method, apparatus and device, and a storage medium. The method comprises the steps of obtaining historical data and processing the historical data to obtain a training data set; performing curve fitting and wavelet decomposition on the training data set to obtain a feature data set; respectively training the gradient boosting decision tree and the gating circulation unit network model to obtain prediction results; and selecting an integration strategy combination prediction result according to the sample number to obtain final prediction power. According to the method, the feature data set fusing the physical law and the multi-scale time sequence features is constructed through curve fitting and wavelet decomposition, the gradient boosting decision tree and the gating circulation unit network are trained in a cooperative mode, and finally the power prediction result is obtained according to the data scale self-adaptive selection integration strategy. The cross-type integration of machine learning, deep learning and physical models is realized, and the prediction precision can be effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of photovoltaic power generation prediction, and in particular to a photovoltaic power station power prediction method, device, equipment and storage medium. BACKGROUND

[0002] In the field of photovoltaic power generation prediction, existing prediction schemes are mainly divided into two categories: modeling methods based on physical mechanisms and mathematical statistical methods based on historical data. Among them, the physical modeling method is heavily dependent on photovoltaic panel parameters and accurate environmental data, and the model error is large under complex weather conditions; while the mainstream data-driven method (such as a single LSTM model or a LightGBM model) can establish a mapping relationship between input and output through data mining, but there is a technical bottleneck: a single model architecture is difficult to simultaneously consider the capture of complex nonlinear relationship between weather and power and the modeling of time series dynamic evolution law, and the pure data-driven nature of the model makes it easy to produce prediction output that violates the basic physical law that "photovoltaic power generation power increases monotonically with solar irradiance" when the training data is sparse or encounters abnormal weather, resulting in limited prediction accuracy and greatly limiting the engineering application value of the prediction result. SUMMARY

[0003] The present application provides a photovoltaic power station power prediction method, device, equipment and storage medium, which is used to solve the problem of limited prediction accuracy of a single model in related technologies.

[0004] The first aspect of the present application provides a photovoltaic power station power prediction method, which comprises: obtaining historical weather data and historical measured power data of a photovoltaic power station, processing the historical weather data and the historical measured power data to obtain a training data set; obtaining a feature data set by curve fitting and wavelet decomposition on the training data set; training a gradient boosting decision tree model according to the feature data set, and generating a first prediction result of the gradient boosting decision tree model based on feature data of a to-be-predicted date; training a gated recurrent unit network model according to the feature data set, and generating a second prediction result of the gated recurrent unit network model based on the feature data of the to-be-predicted date; combining the first prediction result and the second prediction result according to a corresponding integration strategy selected according to the sample number of the training data set, to obtain a predicted power of the photovoltaic power station.

[0005] Optionally, in the first implementation manner of the first aspect of the present application, the step of processing the historical weather data and the historical measured power data to obtain a training data set comprises: randomly sampling and curve fitting irradiance data in the historical meteorological data and the historical measured power data to generate a power reference curve; determining a predicted power value corresponding to each data point in the historical measured power data according to the power reference curve, and obtaining sample residuals of the data points by subtracting the measured power value of the historical measured power data from the predicted power value; calculating a standard deviation of all the sample residuals, and comparing the sample residuals with a positive and negative threshold range set based on the standard deviation to identify a target data point corresponding to a sample residual exceeding the positive and negative threshold range; removing the target data point from the historical meteorological data and historical measured power data to obtain a training data set.

[0006] Optionally, in a second implementation manner of the first aspect of the present application, the step of obtaining a feature data set by curve fitting and wavelet decomposition on the training data set comprises: performing polynomial function fitting on irradiance data and historical measured power data in the training data set, and imposing a monotonically increasing constraint on the fitting result to generate a monotonically increasing power fitting curve; determining a curve predicted value of each sample in the training data set based on the power fitting curve, and determining a curve fitting residual of each sample according to a measured power value of the historical measured power data and the curve predicted value; performing wavelet decomposition on an irradiance time series of historical meteorological data in the training data set to decompose the irradiance time series into an approximation coefficient reflecting a long-term trend and a detail coefficient reflecting fluctuations of different time scales; obtaining wavelet feature data by calculating a mean and a standard deviation of the approximation coefficient and calculating an energy value and an energy proportion of the detail coefficient; integrating the basic meteorological features, time features, the curve predicted value, the curve fitting residual, and the wavelet feature data in the training data set to obtain a feature data set.

[0007] Optionally, in a third implementation manner of the first aspect of the present application, the step of training a gradient boosting decision tree model according to the feature data set and generating a first prediction result of the gradient boosting decision tree model based on feature data of a to-be-predicted date comprises: dividing the feature data set into a training subset and a validation subset, and determining an hyperparameter configuration with a minimum mean absolute error on the validation subset in a preset hyperparameter space; According to the hyperparameter configuration, iterative decision tree construction is performed on the training subset, and in each iteration process, an iteration residual between a predicted value and an actual value of the training subset is determined, a new decision tree is constructed by fitting the iteration residual, and the gradient boosting decision tree model is obtained until the iteration is completed; The feature data of the to-be-predicted date is input into the gradient boosting decision tree model to obtain a first prediction result.

[0008] Optionally, in a fourth implementation manner of the first aspect of the present application, the step of training the gated recurrent unit network model according to the feature data set and generating a second prediction result of the gated recurrent unit network model based on the feature data of the to-be-predicted date comprises: The feature data set is rearranged in time sequence to construct a time sequence sample set composed of consecutive time step samples; A sequence processing network structure is initialized, and the sequence processing network structure is trained according to the time sequence sample set; wherein the sequence processing network structure comprises at least two gated recurrent unit layers, at least one random dropout layer and at least one fully connected layer, and serves as a gated recurrent unit network model to be trained; The feature data of the to-be-predicted date is input into the trained gated recurrent unit network model to obtain the second prediction result.

[0009] Optionally, in a fifth implementation manner of the first aspect of the present application, the step of combining the first prediction result and the second prediction result to obtain the predicted power of the photovoltaic power station according to the sample number of the training data set and the corresponding integration strategy comprises: The sample number of the training data set is compared with preset first and second number thresholds; wherein the first number threshold is less than the second number threshold; When the total sample number is less than the first number threshold, the first prediction result and the second prediction result are weightedly averaged to obtain the predicted power of the photovoltaic power station; When the total sample number is greater than or equal to the first number threshold and less than the second number threshold, the first prediction result is taken as a basic prediction value, a residual prediction value is obtained through residual learning of the gated recurrent unit network model, and the basic prediction value and the residual prediction value are added to obtain the predicted power of the photovoltaic power station; When the total sample number is greater than or equal to the second number threshold, the first prediction result, the second prediction result and a third prediction result generated by the power fitting curve are taken as meta-features to train a meta-learner, and the meta-learner outputs the predicted power of the photovoltaic power station.

[0010] Optionally, in a sixth implementation form of the first aspect of the application, the method further comprises: obtaining a time sequence of the solar elevation angle and the solar azimuth angle of the to-be-predicted date, and combining the installed capacity and the geographic location information of the photovoltaic power station to calculate a time sequence of the theoretical maximum power generation of the to-be-predicted date; comparing the time sequence of the predicted power of the photovoltaic power station with the time sequence of the theoretical maximum power generation point by point, and identifying an abnormal prediction point corresponding to the time point at which the predicted power of the predicted power time sequence exceeds the theoretical maximum power generation; replacing the predicted power of the abnormal prediction point with the theoretical maximum power generation to generate a final power prediction sequence.

[0011] The second aspect of the application provides a photovoltaic power station power prediction device, which is used to implement the photovoltaic power station power prediction method. The photovoltaic power station power prediction device comprises: an acquisition module, configured to acquire historical meteorological data and historical measured power data of a photovoltaic power station, and process the historical meteorological data and the historical measured power data to obtain a training data set; a feature extraction module, configured to obtain a feature data set by performing curve fitting and wavelet decomposition on the training data set; a first prediction module, configured to train a gradient boosting decision tree model according to the feature data set, and generate a first prediction result of the gradient boosting decision tree model based on feature data of a to-be-predicted date; a second prediction module, configured to train a gated recurrent unit network model according to the feature data set, and generate a second prediction result of the gated recurrent unit network model based on the feature data of the to-be-predicted date; a processing module, configured to select a corresponding integration strategy according to the sample number of the training data set, combine the first prediction result and the second prediction result to obtain a predicted power of the photovoltaic power station.

[0012] The third aspect of the embodiment of the application provides an electronic device, comprising 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, each step of the photovoltaic power station power prediction method provided in the first aspect of the embodiment of the application is implemented.

[0013] The fourth aspect of the embodiment of the application provides a computer readable storage medium, which stores a computer program, and when the computer program is executed by a processor, each step of the photovoltaic power station power prediction method provided in the first aspect of the embodiment of the application is implemented.

[0014] In summary, according to the photovoltaic power station power prediction method, device, equipment and storage medium provided by the scheme, historical data is acquired and processed to obtain a training data set; curve fitting and wavelet decomposition are performed on the training data set to obtain a feature data set; a gradient boosting decision tree and a gated recurrent unit network model are trained respectively to obtain a prediction result; an integration strategy is selected according to the number of samples to combine the prediction results, and a final predicted power is obtained. The curve fitting and wavelet decomposition construct a feature data set that fuses physical laws and multi-scale time sequence characteristics, and the gradient boosting decision tree and the gated recurrent unit network are trained cooperatively, and finally the integration strategy is selected adaptively according to the data scale to obtain the power prediction result, realizing cross-type integration of machine learning, deep learning and physical models, and effectively improving the prediction accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 A flowchart of a photovoltaic power station power prediction method provided by an embodiment of the application is shown in the figure. Figure 2 A photovoltaic data cleaning and power reference curve analysis diagram provided by an embodiment of the application is shown in the figure. Figure 3 A monotonic polynomial curve fitting diagram provided by an embodiment of the application is shown in the figure. Figure 4 A training process diagram of a gradient boosting decision tree model provided by an embodiment of the application is shown in the figure. Figure 5 A prediction result diagram of a gradient boosting decision tree model provided by an embodiment of the application is shown in the figure. Figure 6 A training process diagram of a gated recurrent unit network model provided by an embodiment of the application is shown in the figure. Figure 7 A prediction result diagram of a gated recurrent unit network model provided by an embodiment of the application is shown in the figure. Figure 8 A program module diagram of a photovoltaic power station power prediction device provided by an embodiment of the application is shown in the figure. Figure 9 A structure diagram of an electronic device provided by an embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0016] In order to make the purposes, characteristics and advantages of the application more obvious and easy to understand, the technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the application.

[0017] To solve the problem that the prediction accuracy of a single model is limited in the related art, an embodiment of the present application provides a photovoltaic power station power prediction method, which comprises the following steps Figure 1 A flowchart of the photovoltaic power station power prediction method provided by the embodiment is shown in the figure, and the method comprises the following steps: In step 110, historical meteorological data and historical measured power data of the photovoltaic power station are acquired, and the historical meteorological data and the historical measured power data are processed to obtain a training data set.

[0018] Specifically, when the historical meteorological data and the historical measured power data required for photovoltaic power station power prediction are processed, the basic meteorological variables such as irradiance, temperature, and wind speed and the measured output power at the corresponding time point are mainly used. First, the original data from the sensor and the metering device are time-aligned and standardized in format, and then the obviously incorrect negative power, the super-capacity value or the distorted data of the measuring point are removed by cleaning, and then the abnormal samples are identified and removed in combination with the physical correlation and statistical rules, so that the original observation sequence is converted into training data with unified structure and reliable quality, and then a historical database reflecting the state of the photovoltaic system is established.

[0019] In step 120, a characteristic data set is obtained by curve fitting and wavelet decomposition on the training data set.

[0020] Specifically, when the training data is curve fitted and wavelet decomposed, a fitting curve of photovoltaic output is constructed based on the inherent physical relationship between irradiance and measured power, a polynomial form is used and a monotonic increasing constraint is applied, so that the fitting result conforms to the physical property that the power does not decrease when the irradiance increases, and the difference between the fitting value and the measured value is calculated to extract the characteristics of the system deviating from the ideal condition; then the irradiance time series is decomposed into an approximate component reflecting the slow change trend of the sunshine and a multi-scale detail component reflecting the disturbance of the cloud cluster by wavelet decomposition, and further energy distribution and statistical description are extracted, so that the input information covers both the overall change and the short-term fluctuation characteristics, thereby improving the model distinguishability.

[0021] In step 130, a gradient boosting decision tree model is trained according to the characteristic data set, and a first prediction result of the gradient boosting decision tree model is generated based on the characteristic data of the date to be predicted.

[0022] Specifically, the gradient boosting decision tree model is trained based on the characteristic data, the characteristic data is associated with the target power, the tree model structure based on residual iterative learning is used, the error of the previous round is fitted by each round of generated tree, so as to gradually improve the overall prediction accuracy, and the optimal tree depth, learning rate and leaf node number are determined by using the hyperparameter search mechanism, so that the model has high efficiency and robustness in complex nonlinear relationship mining; after the training is completed, the characteristics of the date to be predicted are input into the model, and the predicted power based on the gradient boosting mechanism is generated.

[0023] Step 140, training a gated recurrent unit network model according to the feature data set, and generating a second prediction result of the gated recurrent unit network model based on the feature data of the to-be-predicted date.

[0024] Specifically, the gated recurrent unit network is trained based on the same feature data, the time series samples are input into the multi-layer gated recurrent structure in time sequence, the network extracts short-term changes, trend changes and lag effects by using the update gate and the reset gate, and captures the dependence relationship across multiple time steps in the sequence expansion process, and then combines the random inactivation and the full connection layer to form a deep structure capable of processing nonlinear dynamic characteristics, so that the model obtains the time sequence sensitivity to the cloud movement speed, the irradiation fluctuation and the temperature disturbance, and then outputs another set of prediction results based on the time sequence learning mechanism after providing the to-be-predicted date continuous input.

[0025] Step 150, selecting a corresponding integration strategy according to the number of samples of the training data set to combine the first prediction result and the second prediction result, and obtaining the predicted power of the photovoltaic power station.

[0026] Specifically, after obtaining the outputs of the two types of models, an appropriate integration method is selected according to the size of the training data: in the case of small amount of data, weighted combination is used to avoid overfitting caused by complex structure, in the case of medium scale, the residual of the gradient boosting model is learned by using the sequence network to improve the overall accuracy, and in the case of sufficient amount of data, the meta-learner is trained by using the meta-features containing the two types of prediction results and the fitting curve prediction values, so that the multi-model information is fused in a higher layer manner, thereby generating more stable final prediction power, and forming a complete short-term power prediction process of the photovoltaic power station.

[0027] In an optional embodiment of the present embodiment, the step of processing the historical meteorological data and the historical measured power data to obtain the training data set comprises: randomly sampling and curve fitting the irradiance data in the historical meteorological data and the historical measured power data to generate a power reference curve; determining the predicted power value corresponding to each data point in the historical measured power data according to the power reference curve, and subtracting the measured power value from the predicted power value to obtain the sample residual of each data point; calculating the standard deviation of all sample residuals, and comparing the sample residuals with the positive and negative threshold ranges set based on the standard deviation to identify the target data points corresponding to the sample residuals exceeding the positive and negative threshold ranges; and removing the target data points from the historical meteorological data and the historical measured power data to obtain the training data set.

[0028] In the present embodiment, as Figure 2The photovoltaic data cleaning and power reference curve analysis diagram is shown. The RANSAC (Random Sample Consensus) algorithm is applied to fit the historical irradiance and measured power data. RANSAC identifies and excludes outliers that do not conform to the main data support by repeatedly randomly sampling sub-samples from observation points and estimating model parameters based on the sub-samples, thereby obtaining a robust irradiance-power reference curve. After the power reference curve is determined, the irradiance of all historical observation points is input into the curve, and the predicted power value corresponding to each data point is obtained. The difference between the predicted power value and the historical measured power value is obtained, i.e., the sample residual is formed. The meaning of the sample residual is the deviation between the actual output at this moment and the theoretical output under ideal lighting conditions. If a certain observation point is affected by shading, inverter limitation, sensor drift or partial component failure, the residual will significantly increase. For example, when the irradiance is in strong light conditions but the actual output is lower than the theoretical value due to power station power limitation, the residual will show a significant negative deviation. Conversely, if the sensor error causes the power to be abnormally high, a larger positive deviation will occur. Since the reference curve is subject to monotonic constraints, the residual distribution can directly reflect abnormal conditions, making the difference between abnormal points and normal points more prominent. After obtaining all residuals, in order to determine which observation points should be excluded from the training set, the standard deviation of all residuals needs to be calculated. The standard deviation can represent the dispersion of residuals in the overall distribution. When the residual fluctuation is close to the normal physical response range, the value will be smaller, and the abnormal point will significantly raise its position in the distribution due to the relatively large deviation. By multiplying the standard deviation by a fixed coefficient to establish a positive and negative threshold interval, for example, taking 1.5 times the standard deviation in each direction as the limit, a range that is relatively tolerant of normal fluctuations but has the ability to identify abnormal points can be formed. If the residual of a certain point exceeds this range, it means that the point is inconsistent with the overall law and is a target data point. After the target data point is identified, it is removed from the historical meteorological data and historical power data to generate a training data set that does not contain abnormal samples. The removed data points will not participate in subsequent model training, so that the modeling process is based on more realistic physical response and more stable power output law, thereby improving the reliability and generalization ability of subsequent feature engineering, gradient boosting model training and sequence neural network learning.

[0029] In an optional implementation of the embodiment, the step of obtaining the feature dataset by curve fitting and wavelet decomposition on the training dataset comprises: performing polynomial function fitting on the irradiance data and the historical measured power data in the training dataset, and imposing a monotonically increasing constraint on the fitting result to generate a monotonically increasing power fitting curve; determining a curve prediction value of each sample in the training dataset based on the power fitting curve, and determining a curve fitting residual of each sample according to a measured power value and the curve prediction value of the historical measured power data; performing wavelet decomposition on the irradiance time series of the historical meteorological data in the training dataset, and decomposing the irradiance time series into an approximation coefficient reflecting a long-term trend and a detail coefficient reflecting fluctuations of different time scales; obtaining wavelet feature data by calculating the mean and standard deviation of the approximation coefficient and calculating the energy value and energy proportion of the detail coefficient; and integrating the basic meteorological features, time features, curve prediction values, curve fitting residuals, and wavelet feature data in the training dataset to obtain the feature dataset.

[0030] In the embodiment, a mathematical mapping between irradiance and photovoltaic module output power is first established, so that the nonlinear response relationship between the two is expressed parametrically. A polynomial function takes irradiance as the independent variable, and is composed of a quadratic or cubic expression through a set of coefficients to be solved, which can describe the power growth amplitude brought by light enhancement and the slow rising characteristics in the low irradiance interval, so that the actual photovoltaic module working curve is effectively approximated. The introduction of a monotonically increasing constraint in the solving process makes the derivative of the fitting function not appear negative, thereby maintaining the physical property that "irradiance increases and power does not decrease". The monotonic constraint can be realized by order-preserving regression, which is a kind of optimization method that forces the output sequence to satisfy monotonicity when solving the regression function. By adjusting the fitting value or slope of each interval, local reverse fluctuations caused by random disturbances are reduced, so that the entire curve presents a continuously rising form. For example, Figure 3The single monotonic polynomial curve fitting is shown in the schematic diagram. After the power fitting curve is determined, the irradiance value of each training sample is input into the curve to obtain the curve prediction value. Then, the curve fitting residual is defined according to the difference between the curve prediction value and the historical measured power value, so that each data point has a clear representation in terms of power deviation. If a certain observation point is temporarily blocked by cloud shadow, the measured power will be significantly lower than the curve prediction value, and the residual will show a large negative deviation. If a point appears sensor drift leading to high power, the residual will show a large positive deviation. After completing the polynomial fitting part, the irradiance time series is decomposed by wavelet to split the irradiance change process into a multi-level time-frequency structure. Wavelet decomposition is a mathematical method that integrates time positioning and frequency decomposition capability. By convolving the original time series with a set of wavelet functions with stretching and shifting capability, long-period gentle changes can be separated into approximation coefficients, and rapid changes, jumps and local disturbances can be separated into multiple detail coefficients. The approximation coefficients represent the overall trend of the irradiance over time, such as the gradual increase from sunrise to noon and the gradual decline in the afternoon. The detail coefficients capture short-term cloud passage, intensity shock, and rapid processes such as steep rise and steep fall at different scales. For example, when a large-scale cloud passes, it will show strong fluctuations in the lower frequency detail coefficients, while the higher frequency detail coefficients will capture slight cloud shadow disturbances. By averaging and calculating the standard deviation of the approximation coefficients, the overall light level and stability of the irradiance can be described. By calculating the energy value and energy proportion of the detail coefficients, the contribution of disturbances at different frequencies to the irradiance change can be quantified, allowing the multi-scale meteorological structure to be digitally expressed. After the curve prediction value, curve fitting residual and wavelet features are generated, they are integrated with the basic meteorological features and time features in the training data set, so that each training sample contains irradiance, temperature, wind speed, time position, ideal response power, deviation degree and multi-scale structure from wavelet analysis, so that the prediction model can utilize physical constraints, time series patterns and light disturbance features simultaneously.

[0031] In an optional implementation of the embodiment, the step of training the gradient boosting decision tree model according to the feature data set and generating the first prediction result of the gradient boosting decision tree model based on the feature data of the to-be-predicted date comprises: dividing the feature data set into a training subset and a validation subset, and determining a hyperparameter configuration with the minimum mean absolute error on the validation subset in a preset hyperparameter space; performing iterative decision tree construction on the training subset according to the hyperparameter configuration, and determining an iterative residual between the prediction value and the true value of the training subset in each iteration process, constructing a new decision tree by fitting the iterative residual, and obtaining the gradient boosting decision tree model until the iteration is completed; and inputting the feature data of the to-be-predicted date into the gradient boosting decision tree model to obtain the first prediction result.

[0032] In this embodiment, when modeling the feature dataset, it is divided into a training subset and a validation subset. The overall data is divided into two parts in chronological order or randomly, so that the model learns the mapping relationship on the training subset and evaluates the performance on the validation subset, so as to determine whether the model parameters have stability. Since photovoltaic power prediction has strong time correlation, maintaining time continuity during division can reduce future information leakage and make the model closer to the real business environment. After division, the optimal parameter combination needs to be searched in the preset hyperparameter space, which is a set composed of a series of external control quantities of the model, including learning rate, maximum depth of tree, number of leaf nodes of each tree, and subsampling ratio for controlling model complexity. By traversing multiple combinations in the hyperparameter space and using the mean absolute error as the evaluation indicator, the prediction deviation of the model on the validation subset can be quantified, so that the optimal hyperparameter configuration is determined under the goal of minimizing the error. For example, if the learning rate is set too large, the model will be too sensitive to the single-step residual signal in training and will oscillate, and if the learning rate is too small, it may lead to slow convergence, so the average absolute error of the validation set will increase significantly, so the feedback from the validation set can accurately filter out a more suitable learning rate interval. After obtaining the optimal hyperparameter configuration, the training subset is input into the gradient boosting framework and the model composed of multiple decision trees is constructed. Gradient boosting is a kind of ensemble learning method based on "step-by-step approximation to the true target value", which improves the overall performance by combining multiple weak learners in sequence, where the weak learner usually adopts the decision tree structure. During the construction process, a set of preliminary prediction values will be generated from the initial model, which can be simply taken as the average of the target values of the training set, so that the subsequent update process has a clear starting point. Then, the first round of iteration residual is calculated according to the difference between the true value and the initial prediction value. Iteration residual refers to the offset between the true target and the current model prediction. The core idea of gradient boosting is to gradually fit this offset to make the prediction approach the true state. The iteration residual is input into the weak learner as a new training target to construct a decision tree that reflects the residual structure, and the learning rate is used to control the influence of the tree on the overall model, so as to avoid large-scale updates that cause performance instability. If the residual shows different feature intervals in a certain iteration, for example, the residual is smaller in sunny high irradiance segments and larger in cloudy disturbance intervals, then the newly constructed tree will generate more division nodes in the cloudy disturbance interval to enhance the fitting ability, so that the model can better capture the impact of irradiance fluctuations. The prediction generated by the new tree updates the overall model, making the overall prediction further approach the true value. The above iteration process continues until the specified number of iterations is reached or the validation set error stops decreasing, at which time the complete gradient boosting decision tree model is obtained. For example, Figure 4The training process of the gradient boosting decision tree model is shown in the figure. After the model is trained, the feature data of the date to be predicted is input into the model, so that the model uses the previously learned structure to infer the future irradiance-power relationship. Since the model has learned the comprehensive rules based on the basic meteorological features, multi-scale fluctuation features and power deviation features in the training stage, it can automatically select the appropriate decision path according to the input features in the prediction process. For example, when the input irradiance features show obvious high-frequency fluctuations, the model will use the tree branches corresponding to the high-frequency fluctuations formed in the training, so as to obtain a prediction result that is more sensitive to rapid changes in cloud clusters. When the input features show a smooth rising and falling trend, the model will go to the branch representing the clear sky trend, so that the prediction result is more consistent with the photovoltaic behavior in this period. Through this mechanism, the gradient boosting model can flexibly adjust the prediction path according to the input conditions, and finally output the first prediction result as shown in the figure, so that the photovoltaic power station power prediction has stronger adaptability and accuracy. In the reasoning process of the model, a consistent logical structure is formed in the training stage, so as to guarantee the prediction quality and reliability. Figure 5 The training process of the gradient boosting decision tree model is shown in the figure. After the model is trained, the feature data of the date to be predicted is input into the model, so that the model uses the previously learned structure to infer the future irradiance-power relationship. Since the model has learned the comprehensive rules based on the basic meteorological features, multi-scale fluctuation features and power deviation features in the training stage, it can automatically select the appropriate decision path according to the input features in the prediction process. For example, when the input irradiance features show obvious high-frequency fluctuations, the model will use the tree branches corresponding to the high-frequency fluctuations formed in the training, so as to obtain a prediction result that is more sensitive to rapid changes in cloud clusters. When the input features show a smooth rising and falling trend, the model will go to the branch representing the clear sky trend, so that the prediction result is more consistent with the photovoltaic behavior in this period. Through this mechanism, the gradient boosting model can flexibly adjust the prediction path according to the input conditions, and finally output the first prediction result as shown in the figure, so that the photovoltaic power station power prediction has stronger adaptability and accuracy. In the reasoning process of the model, a consistent logical structure is formed in the training stage, so as to guarantee the prediction quality and reliability.

[0033] In an optional implementation of the embodiment, the step of training the gated recurrent unit network model according to the feature data set and generating a second prediction result of the gated recurrent unit network model based on the feature data of the date to be predicted comprises: rearranging the feature data set in chronological order to construct a time sequence sample set composed of consecutive time step samples; initializing a sequence processing network structure and training the sequence processing network structure according to the time sequence sample set; wherein the sequence processing network structure comprises at least two gated recurrent unit layers, at least one random dropout layer and at least one fully connected layer, serving as the gated recurrent unit network model to be trained; inputting the feature data of the date to be predicted into the trained gated recurrent unit network model to obtain the second prediction result.

[0034] In this embodiment, when constructing the sequence network model based on the gated recurrent unit using the feature dataset, first, all feature samples are rearranged in chronological order, so that the originally single-point stored feature structure is converted into a continuous sequence form that can reflect the chronological relationship, thereby enabling the trend change, short-term disturbance and hysteresis effect of the photovoltaic power evolution over time to be expressed at the data input end. In the sequence construction process, by combining several adjacent time points into a continuous window, each window can serve as a complete time sequence sample, which contains multiple-dimensional variables such as basic meteorological features, curve prediction values, curve fitting residuals and wavelet features, so that the model can capture physical trends, deviation degrees and multi-scale disturbance structures at the same time 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 periodic behavior of irradiance and power within a day, thereby providing the gated recurrent unit with a sample structure with sufficient context information. After completing the sequence construction, the sequence processing network structure is initialized and a deep network containing multiple layers of gated recurrent units, random inactivation layers and fully connected layers is established. The gated recurrent unit is a type of recurrent neural structure for processing time series, in which the update gate and the reset gate are two core components. The update gate is responsible for controlling the extent to which the current unit state inherits the hidden state of the previous time step, and the reset gate is responsible for controlling how the current input and historical state are fused, thereby enabling the network to capture long-term dependencies and short-term changes. By stacking at least two layers of gated recurrent units, the network can identify local rapid fluctuations at a lower level and identify overall trends and slow disturbances at a higher level, so that the model can distinguish between the responses of the photovoltaic system when it is affected by high-frequency cloud shadows and when it is affected by low-frequency weather changes. The random inactivation layer is a structure that randomly masks part of the neurons during the training stage. By randomly setting some activation values to zero during forward propagation, the network avoids over-reliance on specific dimensional features, thereby reducing sensitivity and enhancing model generalization ability. For example, after some wavelet energy features are randomly inactivated in a certain training batch, the model still needs to learn the time series relationship from the remaining features. This reinforcement learning method can make the model perform more stably when it is deployed to face feature disturbances. The fully connected layer is located at the end of the network, and its role is to integrate the time dimension information output by the gated recurrent unit into a single-step prediction. By linear combination and nonlinear activation, the time series features learned by the multi-layer hidden structure are mapped to the power prediction value. For example, Figure 6The training process of the Gated Recurrent Unit network model is shown. After the sequence structure and network architecture are initialized, the entire model is trained through a set of time series samples, and the network parameters are optimized in the continuous forward propagation and back propagation, so that the loss function gradually decreases. During training, the Gated Recurrent Unit processes the feature input step by step, combines the hidden state of the previous moment with the features of the current moment, adjusts the historical memory weight through the update gate, and controls the proportion of new information injection through the reset gate, so that the internal state of each moment has a comprehensive expression of recent fluctuations and long-term trends. For example, during the continuous cloudy disturbance phase, the reset gate will enhance the influence of the current input, while in the clear sky stable phase, the update gate will strengthen the inheritance of long-term trends, so that the model can automatically adapt to different weather process time expression methods, and finally form a complete parameter set. With the advancement of training, the network gradually learns the dynamic relationship between irradiance changes, wavelet energy distribution, power deviation structure and time factors, enabling it to extract patterns from historical data and predict future conditions. When the training reaches the convergence condition, the feature data of the date to be predicted is input into the network in the same format as the training phase, so that the network propagates the input information in the time dimension step by step, and outputs the first prediction result in the last fully connected layer. Figure 7 The second prediction result is shown.

[0035] In an optional implementation of the present embodiment, the step of combining the first prediction result and the second prediction result to obtain the predicted power of the photovoltaic power station according to the selected corresponding ensemble strategy based on the number of samples in the training data set comprises: comparing the number of samples in the training data set with a preset first number threshold and a second number threshold; wherein the first number threshold is less than the second number threshold; when the total number of samples is less than the first number threshold, the first prediction result and the second prediction result are weighted and averaged to obtain the predicted power of the photovoltaic power station; when the total number of samples is greater than or equal to the first number threshold and less than the second number threshold, the first prediction result is taken as a basic prediction value, and a residual prediction value is obtained through residual learning of the Gated Recurrent Unit network model, and the basic 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 number threshold, the first prediction result, the second prediction result and a third prediction result generated by the power fitting curve are taken as meta-features to train a meta-learner, and the meta-learner outputs the predicted power of the photovoltaic power station.

[0036] Specifically, in the power prediction of photovoltaic power stations, 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 the preset first quantity threshold and the second quantity threshold, an adaptive prediction strategy can be realized. According to the sample size of the training dataset, the optimal ensemble strategy is dynamically selected, so that the optimal balance between prediction accuracy and model stability can be achieved in three different scenarios of data scarcity, adequacy and abundance. The first quantity threshold is less than the second quantity threshold. When the total number of samples is less than the first quantity threshold, it indicates that the historical information available for model training is limited. At this time, if a complex ensemble model is used, it is easy to cause overfitting, that is, the model overfits the noise in the training data and loses the generalization ability. Therefore, in this data scarce scenario, a weighted average strategy with high computational efficiency and simple structure is adopted. Specifically, the optimal weight coefficients of the first prediction result of the gradient boosting decision tree model and the second prediction result of the gated recurrent unit network model are determined on a validation set independent of the training process through grid search or optimization algorithm, with the constraint that the sum of the weights is 1. The final predicted power is the linear superposition of the weighted first prediction result and the second prediction result. The advantage of this strategy is that it integrates the prediction tendencies of the two heterogeneous models with the lowest model complexity, effectively avoiding the problem of unstable model performance caused by insufficient data. When the total number of samples reaches or exceeds the first quantity threshold but is still less than the second quantity threshold, it means that the data volume is sufficient to support more sophisticated model collaboration. At this time, the residual learning ensemble strategy is enabled. 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 goal, i.e., instead of directly predicting power, it is used to fit the residual error between the base prediction value and the true power value, which represents the complex time series fluctuations and nonlinear patterns that the gradient boosting decision tree model cannot explain. The gated recurrent unit network can effectively capture the long-term dependencies in time series through its internal gating mechanism (update gate and reset gate), thus accurately learning these residual patterns. The final predicted power is composed of the base prediction value and the residual prediction value output by the gated recurrent unit network, forming a cascading and complementary prediction process. When the sample size further expands to reach or exceed the second quantity threshold, it indicates that there is a large amount of diverse historical data, which is sufficient to train a more complex but potentially higher 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 calculated directly from the power fitting curve (a benchmark curve that strictly satisfies the physical law that power increases monotonically with irradiance, generated by polynomial fitting and order-preserving regression) are jointly used as meta-features.These meta-features combine different perspectives and predictive wisdom from machine learning models, deep learning models, and physical constraint models. The second layer introduces a meta-learner, usually a relatively simple shallow neural network, whose input is the meta-feature vector composed of the three meta-predictions. The meta-learner learns on the training set how to optimally weight and combine these meta-features, and its output is the final, further refined photovoltaic power station prediction. This stacking ensemble strategy automatically uncovers the complex interactions and dependencies among the prediction results of different base models through the meta-learner, thereby maximizing the prediction performance under sufficient data conditions.

[0037] In an optional implementation of the embodiment, the time series of the solar elevation angle and the solar azimuth angle of the to-be-predicted date are obtained, and the time series of the theoretical maximum power generation of the to-be-predicted date is calculated by combining the installed capacity and the geographic location information of the photovoltaic power station; the time series of the predicted power of the photovoltaic power station and the time series of the theoretical maximum power generation are compared point by point, and the abnormal prediction point corresponding to the predicted power of the time series of the predicted power exceeding the theoretical maximum power generation at the same time point is identified; the predicted power of the abnormal prediction point is replaced by the theoretical maximum power generation to generate the final power prediction sequence.

[0038] In this embodiment, the solar elevation angle is defined as the angle between the solar ray and the horizontal plane, with a value range of -90 degrees to +90 degrees, reaching the maximum at noon; the solar azimuth angle represents the angle between the projection of the sun on the horizontal plane and the north direction, with a value range of 0 degrees to 360 degrees. Based on the specific latitude and longitude coordinates of the photovoltaic power station and the annual and time sequence information of the date to be predicted, the astronomical equation containing parameters such as declination angle, hour angle, and solar time is iteratively solved to finally generate a time sequence of solar elevation angle and solar azimuth angle that completely matches the prediction time resolution. After obtaining the accurate celestial orbit data, combined with the installed capacity of the photovoltaic power station, i.e. the rated maximum power value of the power station, the theoretical maximum power time sequence is calculated through the photovoltaic power generation physical model. This calculation process needs to consider factors such as the geometric relationship of the sun's position, the inclination and azimuth configuration of the photovoltaic panel, and the atmospheric optical quality, among which the incident angle correction coefficient needs to be introduced in particular to quantify the influence of the angle between the solar ray and the normal line of the photovoltaic panel on the energy receiving efficiency. When the solar ray is perpendicular to the photovoltaic panel, the receiving efficiency is the highest, and as the incident angle increases, the cosine energy loss will occur. The calculation formula of the theoretical maximum power can be expressed as: the theoretical power is equal to the product of the installed capacity and the normal direct irradiance, multiplied by the composite trigonometric function composed of the sine function of the solar elevation angle and the cosine function of the azimuth angle difference, and finally the temperature correction coefficient and the system efficiency coefficient are introduced for calibration. Through this series of calculations, the theoretical power upper limit curve that conforms to the physical characteristics of the power station and the local astronomical law can be obtained. After completing the establishment of the theoretical benchmark, the original predicted power time sequence output by the data-driven model is compared with the theoretical maximum power time sequence at each time point, and the abnormal data points whose predicted power values exceed the theoretical maximum power values at the same time point are accurately identified. The generation of such abnormal points is often due to the fact that the machine learning model has learned the data pattern containing measurement errors or abnormal operating states in the training process, resulting in excessive prediction that violates the physical law under certain weather conditions. For the identified abnormal prediction points, the system starts the automatic correction mechanism to replace the predicted power value with the theoretical maximum power value at the corresponding time. This replacement operation maintains the integrity of the prediction results of non-abnormal points, only intervenes in the data points that violate the physical law, and finally generates the final power prediction sequence that retains the intelligent prediction characteristics of the data-driven model and strictly follows the physical law constraints. This correction mechanism effectively solves the power overproduction prediction problem that may be caused by pure data-driven methods, for example, in the rapid change process of cloudy weather turning sunny, the model may produce power prediction exceeding the physical limit of the power station due to learning historical abnormal data, and through this physical constraint correction, it can ensure that the prediction result is always within the reasonable physical boundary.

[0039] According to the photovoltaic power station power prediction method provided in the scheme, historical data is acquired and processed to obtain a training data set; curve fitting and wavelet decomposition are performed on the training data set to obtain a feature data set; a gradient boosting decision tree and a gated recurrent unit network model are trained respectively to obtain a prediction result; an integration strategy is selected according to the number of samples to combine the prediction results, and a final predicted power is obtained. In the scheme, the curve fitting and wavelet decomposition are used to construct a feature data set that fuses physical laws and multi-scale time sequence characteristics, and the gradient boosting decision tree and the gated recurrent unit network are trained cooperatively, and finally an integration strategy is selected adaptively according to the data scale to obtain a power prediction result, realizing cross-type integration of machine learning, deep learning and physical models, and effectively improving the prediction accuracy.

[0040] Figure 2 A photovoltaic power station power prediction device is provided in the embodiments of the present application, and the photovoltaic power station power prediction device can be used to implement the photovoltaic power station power prediction method in the foregoing embodiments. As shown in the device, Figure 2 The device mainly includes: An acquisition module 10 is configured to acquire historical meteorological data and historical measured power data of a photovoltaic power station, and process the historical meteorological data and the historical measured power data to obtain a training data set; A feature extraction module 20 is configured to perform curve fitting and wavelet decomposition on the training data set to obtain a feature data set; A first prediction module 30 is configured to train a gradient boosting decision tree model according to the feature data set, and generate a first prediction result of the gradient boosting decision tree model based on feature data of a to-be-predicted date; A second prediction module 40 is configured to train a gated recurrent unit network model according to the feature data set, and generate a second prediction result of the gated recurrent unit network model based on the feature data of the to-be-predicted date; A processing module 50 is configured to select a corresponding integration strategy according to the number of samples of the training data set, combine the first prediction result and the second prediction result, and obtain a predicted power of the photovoltaic power station.

[0041] In an optional implementation of the embodiments, the acquisition module is specifically configured to: perform random sampling and curve fitting on the irradiance data in the historical meteorological data and the historical measured power data to generate a power reference curve; determine a predicted power value corresponding to each data point in the historical measured power data according to the power reference curve, and obtain a sample residual error of each data point by subtracting the predicted power value from the measured power value of the historical measured power data; calculate a standard deviation of all sample residual errors, and compare the sample residual errors with a positive and negative threshold range set based on the standard deviation to identify target data points corresponding to sample residual errors that exceed the positive and negative threshold range; and remove the target data points from the historical meteorological data and the historical measured power data to obtain the training data set.

[0042] In an optional implementation of the embodiment, the feature extraction module is specifically configured to: perform polynomial function fitting on the irradiance data and the historical measured power data in the training data set, and impose a monotonically increasing constraint on the fitting result to generate a monotonically increasing power fitting curve; determine a curve prediction value of each sample in the training data set based on the power fitting curve, and determine a curve fitting residual of each sample according to a measured power value and the curve prediction value of the historical measured power data; perform wavelet decomposition on the irradiance time series of the historical meteorological data in the training data set, to decompose the irradiance time series into an approximation coefficient reflecting a long-term trend and a detail coefficient reflecting fluctuations of different time scales; obtain wavelet feature data by calculating the mean and standard deviation of the approximation coefficient, and calculating the energy value and the energy proportion of the detail coefficient, respectively; and integrate the basic meteorological features, the time features, the curve prediction value and the curve fitting residual, and the wavelet feature data in the training data set to obtain the feature data set.

[0043] In an optional implementation of the embodiment, the first prediction module is specifically configured to: divide the feature data set into a training subset and a validation subset, and determine an hyperparameter configuration with the minimum mean absolute error on the validation subset in a preset hyperparameter space; perform iterative decision tree construction on the training subset according to the hyperparameter configuration, and determine an iterative residual between the predicted value and the true value of the training subset in each iteration process, construct a new decision tree by fitting the iterative residual, and obtain a gradient boosting decision tree model until the iteration is completed; and input the feature data of the date to be predicted into the gradient boosting decision tree model to obtain the first prediction result.

[0044] In an optional implementation of the embodiment, the second prediction module is specifically configured to: rearrange the feature data set in chronological order to construct a time series sample set composed of consecutive time step samples; initialize a sequence processing network structure, and train the sequence processing network structure according to the time series sample set; wherein the sequence processing network structure includes at least two gate recurrent unit layers, at least one random dropout layer and at least one fully connected layer, as a gate recurrent unit network model to be trained; and input the feature data of the date to be predicted into the trained gate recurrent unit network model to obtain the second prediction result.

[0045] In an optional implementation of the embodiment, the processing module is specifically configured to: compare the sample quantity of the training data set with preset first and second quantity thresholds; the first quantity threshold is smaller than the second quantity threshold; when the total sample quantity is smaller 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; when the total sample quantity is greater than or equal to the first quantity threshold and smaller than the second quantity threshold, the first prediction result is taken as a basic prediction value, and a residual prediction value is obtained through residual learning of the gated recurrent unit network model, and the basic prediction value and the residual prediction value are added to obtain the predicted power of the photovoltaic power station; when the total sample quantity is greater than or equal to the second quantity threshold, the first prediction result, the second prediction result, and a third prediction result generated by the power fitting curve are taken as meta-features to train a meta-learner, and the meta-learner is used to output the predicted power of the photovoltaic power station.

[0046] In an optional implementation of the embodiment, the processing module is further configured to: obtain the time sequence of the solar elevation angle and the solar azimuth angle of the to-be-predicted date, and combine the installed capacity and the geographic location information of the photovoltaic power station to calculate the time sequence of the theoretical maximum power generation of the to-be-predicted date; compare the time sequence of the predicted power of the photovoltaic power station with the time sequence of the theoretical maximum power generation point by point, and identify an abnormal prediction point corresponding to the predicted power of the predicted power time sequence that exceeds the theoretical maximum power generation at the same time point; replace the predicted power of the abnormal prediction point with the theoretical maximum power generation to generate a final power prediction sequence.

[0047] According to the photovoltaic power station power prediction device provided in the scheme, historical data is obtained and processed to obtain a training data set; curve fitting and wavelet decomposition are performed on the training data set to obtain a feature data set; a gradient boosting decision tree and a gated recurrent unit network model are trained to obtain prediction results; and an integration strategy is selected according to the sample quantity to combine the prediction results, so as to obtain a final predicted power. In the scheme, curve fitting and wavelet decomposition are used to construct a feature data set that fuses physical laws and multi-scale time sequence features, and a gradient boosting decision tree and a gated recurrent unit network are trained in cooperation, and finally an integration strategy is selected according to the data scale to obtain a power prediction result, so that cross-type integration of machine learning, deep learning, and a physical model is realized, and the prediction accuracy can be effectively improved.

[0048] According to the photovoltaic power station power prediction device provided in the scheme, Figure 9 An electronic device is provided for the embodiment. The electronic device can be used to implement the photovoltaic power station power prediction method in the foregoing embodiments, and mainly includes: The memory 901, the processor 902 and the computer program 903 stored on the memory 901 and executable on the processor 902 are connected by communication. When the processor 902 executes the computer program 903, the photovoltaic power plant power prediction method in the foregoing embodiments is implemented. The number of processors can be one or more.

[0049] The memory 901 can be a high-speed random access memory (RAM) or a non-volatile memory such as a disk memory. The memory 901 is used to store executable program codes, and the processor 902 is coupled to the memory 901.

[0050] Further, the embodiment of the present application also provides a computer readable storage medium, which can be arranged in the electronic device in the foregoing embodiments. The computer readable storage medium can be the memory in the foregoing embodiments. Figure 3

[0051] The computer readable storage medium stores a computer program, and the program is executed by the processor to implement the photovoltaic power plant power prediction method in the foregoing embodiments. Further, the computer readable storage medium can also be a U disk, a mobile hard disk, a read-only memory (ROM), a RAM, a magnetic disk or an optical disk, and various media that can store program codes.

[0052] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system, device and unit described above can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0053] The integrated unit, if realized in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the embodiments of the present application. The foregoing storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0054] ​The above-described embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present 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. 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. 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.

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 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: 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. 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 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.

5. The photovoltaic power plant power prediction method according to claim 4, characterized in that, 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: The feature dataset is rearranged in chronological order to construct a time series sample set consisting of samples from consecutive time steps; 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; 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.

6. The photovoltaic power plant power prediction method according to claim 3, characterized in that, 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: 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. 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. 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. 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.

7. 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.

8. 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.

9. 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 7.

10. 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 7.

Images