Photovoltaic power generation power prediction method, device, equipment, medium and product

Abstract

The embodiment of the invention provides a photovoltaic power generation power prediction method and device, equipment, a medium and a product, and is applied to the technical field of photovoltaic power generation. The method comprises the following steps: acquiring real-time meteorological data and power time sequence data corresponding to the real-time meteorological data; calculating theoretical generated power corresponding to the real-time meteorological data based on a preset physical model; calculating predicted power generation power corresponding to the power time sequence data based on a pre-trained mathematical model; calculating corrected power based on the theoretical generated power and the predicted generated power; calculating a dynamic weight based on the real-time meteorological data and the correction power; and performing weighted calculation based on the dynamic weight, the correction power and the theoretical generation power to obtain target prediction power corresponding to the real-time meteorological data. The method achieves the technical effect of improving the accuracy of photovoltaic power generation power prediction.

Description

Technical Field

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

[0002] The intermittent and fluctuating nature of renewable energy generation presents challenges for grid integration and dispatch. Therefore, accurate power forecasting not only contributes to the stable operation of the power system but also effectively improves the utilization efficiency of renewable energy and reduces the demand for backup power. Furthermore, with the gradual opening of the electricity market, renewable energy power forecasting has a profound impact on market transactions, price formation, and policy formulation. Against this backdrop, developing advanced renewable energy power forecasting technologies, researching advanced power forecasting algorithms, and combining various information such as meteorological data and historical power generation data can provide strong support for the safety and economy of power system operation, and promote the further development and application of renewable energy.

[0003] In existing technologies, photovoltaic power generation prediction methods are mainly divided into two categories: physical model methods and data-driven methods. Physical models are based on the electrical characteristic equations of photovoltaic modules and combine meteorological data to calculate theoretical power generation for prediction; data-driven methods establish statistical models by analyzing historical data to make predictions.

[0004] Because existing prediction methods comprehensively consider the influence of multiple factors, there is a technical problem of low accuracy in photovoltaic power generation prediction. Summary of the Invention

[0005] This application provides a method, apparatus, equipment, medium, and product for predicting photovoltaic power generation, in order to achieve the technical effect of improving the accuracy of photovoltaic power generation prediction.

[0006] In a first aspect, embodiments of this application provide a photovoltaic power generation prediction method, including:

[0007] Acquire real-time meteorological data and the corresponding power time-series data;

[0008] The theoretical power generation capacity corresponding to real-time meteorological data is calculated based on a preset physical model.

[0009] The predicted power generation is calculated based on the power time series data using a pre-trained mathematical model.

[0010] Corrected power is calculated based on theoretical power generation and predicted power generation.

[0011] Dynamic weights are calculated based on real-time meteorological data and corrected power.

[0012] The target predicted power corresponding to the real-time meteorological data is obtained by weighting the dynamic weight, the corrected power, and the theoretical power generation.

[0013] In one possible implementation, dynamic weights are calculated based on real-time meteorological data and corrected power, including:

[0014] Based on real-time meteorological data and corrected power, the confidence scores of the physical model and the mathematical model are calculated.

[0015] Based on preset hyperparameters, the confidence scores of the physical model and the mathematical model are weighted and normalized to obtain dynamic weights.

[0016] In one possible implementation, before acquiring real-time meteorological data and the corresponding power time-series data, the method further includes:

[0017] Acquire historical meteorological data and historical power data;

[0018] Based on historical meteorological data and a pre-set physical model, the theoretical power value and physical parameters are calculated.

[0019] Based on theoretical power values, historical power data are processed and features are extracted to obtain sample training data.

[0020] The initial mathematical model is iteratively distilled and trained based on theoretical power values, physical parameters, and sample training data to obtain the target mathematical model.

[0021] In one possible implementation, historical power data is processed and features are extracted based on theoretical power values ​​to obtain sample training data, including:

[0022] Outlier filtering is performed on historical power data based on theoretical power values ​​to obtain historical power data with outliers removed.

[0023] Missing values ​​are filled in and normalized for historical power data after outliers are removed to obtain standard power data.

[0024] Multi-scale periodic feature extraction is performed based on standard power data to obtain sample training data.

[0025] In one possible implementation, the initial mathematical model is iteratively distilled and trained based on theoretical power values, physical parameters, and sample training data to obtain the target mathematical model, including:

[0026] In each iteration of distillation training, the periodic encoding layer based on the initial mathematical model transforms each multi-scale periodic feature data in the sample training data into an initial hidden state.

[0027] Based on the TimesNet core layer in the initial mathematical model, time-frequency transformation is performed on the initial hidden state and multi-scale periodic features to obtain the time-series feature vector.

[0028] The original predicted output value corresponding to the temporal feature vector is calculated based on the fully connected layer in the initial mathematical model;

[0029] The target predicted output value is obtained by constraining the original predicted output value based on the physical parameters.

[0030] Distillation training is performed based on the theoretical power value and the target predicted output value to obtain the target mathematical model.

[0031] In one possible implementation, distillation training is performed based on the theoretical power value and the target predicted output value to obtain the target mathematical model, including:

[0032] Based on the theoretical power value and the target predicted output value, feature decomposition is performed to obtain the first short-term component, the first daily cycle component, and the first long-term component corresponding to the theoretical power value, and the second short-term component, the second daily cycle component, and the second long-term component corresponding to the target predicted output value.

[0033] Distillation loss is calculated based on the first short-term component, the first daily cycle component, the first long-term component, the second short-term component, the second daily cycle component, and the second long-term component.

[0034] The prediction loss is calculated based on the target prediction output value and the original output value corresponding to the sample training data.

[0035] The total loss is obtained by weighting the physical regularization input, prediction loss and distillation loss corresponding to the preset physical model.

[0036] When the total loss is less than a preset threshold, the mathematical model in the current training phase is determined as the target mathematical model.

[0037] Secondly, embodiments of this application provide a photovoltaic power generation prediction device, comprising:

[0038] The acquisition module is used to acquire real-time meteorological data and the corresponding power time-series data.

[0039] The first processing module is used to calculate the theoretical power generation corresponding to real-time meteorological data based on a preset physical model.

[0040] The second processing module is used to calculate the predicted power generation corresponding to the power time series data based on a pre-trained mathematical model.

[0041] The third processing module is used to calculate the corrected power based on the theoretical power generation and the predicted power generation.

[0042] The fourth processing module is used to calculate dynamic weights based on real-time meteorological data and corrected power.

[0043] The fifth processing module is used to perform weighted calculations based on dynamic weights, corrected power, and theoretical power generation to obtain the target predicted power corresponding to the real-time meteorological data.

[0044] In one possible implementation, the fourth processing module is further configured to:

[0045] Based on real-time meteorological data and corrected power, the confidence scores of the physical model and the mathematical model are calculated.

[0046] Based on preset hyperparameters, the confidence scores of the physical model and the mathematical model are weighted and normalized to obtain dynamic weights.

[0047] In one possible implementation, the device further includes a sixth processing module for:

[0048] Acquire historical meteorological data and historical power data;

[0049] Based on historical meteorological data and a pre-set physical model, the theoretical power value and physical parameters are calculated.

[0050] Based on theoretical power values, historical power data are processed and features are extracted to obtain sample training data.

[0051] The initial mathematical model is iteratively distilled and trained based on theoretical power values, physical parameters, and sample training data to obtain the target mathematical model.

[0052] In one possible implementation, the sixth processing module is further configured to:

[0053] Outlier filtering is performed on historical power data based on theoretical power values ​​to obtain historical power data with outliers removed.

[0054] Missing values ​​are filled in and normalized for historical power data after outliers are removed to obtain standard power data.

[0055] Multi-scale periodic feature extraction is performed based on standard power data to obtain sample training data.

[0056] In one possible implementation, the sixth processing module is further configured to:

[0057] In each iteration of distillation training, the periodic encoding layer based on the initial mathematical model transforms each multi-scale periodic feature data in the sample training data into an initial hidden state.

[0058] Based on the TimesNet core layer in the initial mathematical model, time-frequency transformation is performed on the initial hidden state and multi-scale periodic features to obtain the time-series feature vector.

[0059] The original predicted output value corresponding to the temporal feature vector is calculated based on the fully connected layer in the initial mathematical model;

[0060] The target predicted output value is obtained by constraining the original predicted output value based on the physical parameters.

[0061] Distillation training is performed based on the theoretical power value and the target predicted output value to obtain the target mathematical model.

[0062] In one possible implementation, the sixth processing module is further configured to:

[0063] Based on the theoretical power value and the target predicted output value, feature decomposition is performed to obtain the first short-term component, the first daily cycle component, and the first long-term component corresponding to the theoretical power value, and the second short-term component, the second daily cycle component, and the second long-term component corresponding to the target predicted output value.

[0064] Distillation loss is calculated based on the first short-term component, the first daily cycle component, the first long-term component, the second short-term component, the second daily cycle component, and the second long-term component.

[0065] The prediction loss is calculated based on the target prediction output value and the original output value corresponding to the sample training data.

[0066] The total loss is obtained by weighting the physical regularization input, prediction loss and distillation loss corresponding to the preset physical model.

[0067] When the total loss is less than a preset threshold, the mathematical model in the current training phase is determined as the target mathematical model.

[0068] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor;

[0069] The memory stores the instructions that the computer executes;

[0070] The processor executes computer execution instructions stored in memory, causing the processor to perform the first aspect above and various possible implementations of the first aspect.

[0071] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and various possible implementations thereof.

[0072] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and various possible implementations thereof.

[0073] This application provides a photovoltaic power generation prediction method, apparatus, equipment, medium, and product. The method acquires real-time meteorological data and corresponding power time-series data; processes the real-time meteorological data using a preset physical model to obtain theoretical power generation; processes the power time-series data using a pre-trained mathematical model to obtain predicted power generation; calculates corrected power using the theoretical power generation and predicted power generation; calculates dynamic weights using the corrected power and real-time meteorological data; and performs a weighted calculation on the corrected power and theoretical power generation using the dynamic weights to obtain the target predicted power. Compared with existing technologies, this application uses the predicted power generation calculated by a mathematical model as a correction term, and dynamically corrects the theoretical power generation by calculating the corrected power, thereby obtaining a prediction result that conforms to physical laws and data propagation laws, achieving the technical effect of improving the accuracy of photovoltaic power generation prediction. Attached Figure Description

[0074] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0075] Figure 1 A flowchart illustrating the photovoltaic power generation prediction method provided in this application. Figure 1 ;

[0076] Figure 2 A flowchart illustrating the photovoltaic power generation prediction method provided in this application. Figure 2 ;

[0077] Figure 3 A flowchart illustrating the photovoltaic power generation prediction method provided in this application. Figure 3 ;

[0078] Figure 4 A schematic diagram of the photovoltaic power generation prediction device provided in this application;

[0079] Figure 5 A schematic diagram of the structure of the electronic device provided in this application.

[0080] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0081] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0082] First, let's explain the terms used in this application:

[0083] Time series two-dimensional change modeling for general time series analysis (TimesNet): refers to a time series prediction model based on deep learning that can capture complex patterns in time series and make accurate predictions about the future.

[0084] In existing technologies, photovoltaic power generation prediction methods are mainly divided into two categories: physical model method and data-driven method. The physical model method solves the equations by analyzing the degradation characteristics of photovoltaic modules to obtain the theoretical power generation value; or the data-driven method uses historical data for training to perform time-series prediction and obtain the predicted power generation value.

[0085] However, existing physical models have large prediction errors under complex weather conditions, making it difficult to accurately reflect actual power fluctuations; data-driven methods suffer from poor consistency of physical laws, leading to prediction results that violate the basic operating principles of photovoltaic systems; thus, existing technologies suffer from low accuracy in photovoltaic power generation prediction.

[0086] To address the aforementioned technical problems, this application proposes the following technical concept: Acquire current meteorological data and power time-series data; calculate the theoretical power generation corresponding to the real-time meteorological data based on a physical model; calculate the predicted power generation corresponding to the power time-series data based on a pre-trained mathematical model; calculate the corrected power generation using the theoretical and predicted power generation; and calculate the target predicted power using a weighted calculation method with dynamic weights. Compared with existing technologies, this application uses a mathematical model to dynamically correct the theoretical calculation results of the physical model, thereby obtaining prediction results that conform to physical laws and data propagation laws, achieving the technical effect of improving the accuracy of photovoltaic power generation prediction.

[0087] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0088] Figure 1 A flowchart illustrating the photovoltaic power generation prediction method provided in this application. Figure 1 ,like Figure 1 As shown, the method includes:

[0089] S101. Obtain real-time meteorological data and the corresponding power time-series data.

[0090] In this step, the method for obtaining real-time meteorological data can be as follows:

[0091] Real-time meteorological data can be obtained through meteorological data interfaces; real-time weather information can be obtained through meteorological satellites and radar; meteorological data can be collected in real time through IoT sensors deployed in cities or specific areas; or meteorological data can be collected through mobile applications and websites. Meteorological data includes, but is not limited to: temperature data, humidity data, air pressure data, wind data, and weather data.

[0092] In this step, the power timing data can be obtained in the following ways:

[0093] Power time-series data is obtained by acquiring power data within a preset time period based on sensors or monitoring systems associated with photovoltaic power generation equipment.

[0094] S102. Calculate the theoretical power generation corresponding to real-time meteorological data based on the preset physical model.

[0095] In this step, the theoretical power generation calculated based on the preset physical model is the theoretical photovoltaic power generation corresponding to the current real-time meteorological data. The method for calculating the theoretical power generation is shown in Formula 1:

[0096]

[0097] Formula 1 corresponds to a preset physical model, which is used to calculate the theoretical power generation based on real-time meteorological data. teacher This refers to the calculation function of the pre-defined physical model; G tilt This refers to the irradiance of the inclined plane, T cell This refers to ambient temperature P. phy (t) refers to the theoretical power generation at time t predicted by the preset physical model.

[0098] S103. Calculate the predicted power generation corresponding to the power time series data based on the pre-trained mathematical model.

[0099] Alternatively, one possible way to calculate the predicted power generation is as follows:

[0100] S1032. Based on power time series data, perform data processing and feature extraction to obtain power sequence and periodic features. Combine real-time meteorological data to determine the meteorological change markers corresponding to the power sequence and obtain feature input data.

[0101] S1032. Input the feature input data into the pre-trained mathematical model to obtain the predicted power generation. The calculation method for the predicted power generation is shown in Formula 2:

[0102]

[0103] Among them, P stu (t) refers to the predicted power generation at time t obtained from the mathematical model, Xt refers to the feature input data, and f student This refers to the mathematical model that has been pre-trained.

[0104] S104. Calculate the corrected power based on theoretical power generation and predicted power generation.

[0105] In this step, the corrected power refers to the residual between the theoretical power generation and the predicted power generation.

[0106] For example, when the calculated theoretical power generation is 150KW and the predicted power generation is 145KW, the corresponding correction power is -5KW; when the calculated theoretical power generation is 150KW and the predicted power generation is 165KW, the corresponding correction power is +15KW.

[0107] The purpose of calculating the corrected power in this step is to calculate the subsequent dynamic weights.

[0108] S105. Calculate dynamic weights based on real-time meteorological data and corrected power.

[0109] Alternatively, one possible implementation of dynamic weight calculation is as follows:

[0110] S1051. Based on real-time meteorological data and corrected power, the confidence scores of the physical model and mathematical model are calculated.

[0111] In this step, the confidence level of the physical model is calculated as shown in Formula 3:

[0112]

[0113] Among them, C phy This refers to the confidence level of the physical model, which depends on the irradiance prediction error; G forecast This refers to the predicted irradiance, G. actual This refers to the actual observed irradiance; γ refers to the sensitivity coefficient.

[0114] The confidence level of the mathematical model is calculated as shown in Formula 4:

[0115]

[0116] Among them, Cnn This refers to the confidence level of the mathematical model, where ΔP refers to the correction power. rated This refers to the rated power output of the photovoltaic array.

[0117] S1052. Based on preset hyperparameters, perform weighted normalization calculation on the confidence scores of the physical model and the mathematical model to obtain dynamic weights.

[0118] In this step, the dynamic weights are calculated as shown in Formula 5:

[0119]

[0120] in, This refers to the calculated dynamic weights, C phy This refers to the confidence level of the physical model, C. nn This refers to the confidence level of the mathematical model. α is the sigmoid function; α and β are adjustable preset hyperparameters. This represents the current rate of change in irradiance. The fusion weights are dynamically calculated; when the physical model has high confidence and the weather is stable, the dynamic weights are close to 1; when the mathematical model has high confidence or the weather changes abruptly, the dynamic weights are close to 0.

[0121] S106. Based on dynamic weights, corrected power, and theoretical power generation, a weighted calculation is performed to obtain the target predicted power corresponding to the real-time meteorological data.

[0122] In this step, the weighted calculation method for obtaining the target predicted power is shown in Formula 6:

[0123]

[0124] Among them, P final (t) refers to the target predicted power at time t, and ΔP refers to the corrected power, P phy (t) refers to the theoretical power generation at time t predicted by the preset physical model. This refers to the calculated dynamic weights.

[0125] It should be noted that the better the dynamic weights are calculated, the more the hyperparameters need to be updated periodically. When the update cycle is 15 minutes, the hyperparameters need to be adjusted every 15 minutes based on the recent prediction errors of the preset physical and mathematical models. The adjustment method is shown in Formula 7:

[0126]

[0127] Among them, RMSE phy This refers to the root mean square error (RMSE) of the pre-defined physical model within the most recent time window. nnThis refers to the root mean square error of the mathematical model; sgn refers to the sign function, which returns 1 when the physical error is large and -1 otherwise. This refers to the update step size; α refers to the preset hyperparameter, and both preset hyperparameters can be updated using this formula.

[0128] The photovoltaic power generation prediction method provided in this application involves acquiring real-time meteorological data and corresponding power time-series data, processing the real-time meteorological data using a preset physical model to obtain theoretical power generation, processing the power time-series data using a pre-trained mathematical model to obtain predicted power generation, calculating a corrected power using the theoretical power generation and predicted power generation, calculating dynamic weights using the corrected power and real-time meteorological data, and weighting the corrected power and theoretical power generation using the dynamic weights to obtain the target predicted power. Compared with existing technologies, this application uses the predicted power generation calculated by a mathematical model as a correction term, and dynamically corrects the theoretical power generation by calculating the corrected power, thereby obtaining a prediction result that conforms to physical laws and data propagation laws, achieving the technical effect of improving the accuracy of photovoltaic power generation prediction.

[0129] Figure 2 A flowchart illustrating the photovoltaic power generation prediction method provided in this application. Figure 2 ,like Figure 2 As shown, the method includes:

[0130] S201. Obtain historical meteorological data and historical power data.

[0131] In this step, historical meteorological data and historical power data can be obtained by either retrieving them from the log database corresponding to the photovoltaic array or from the monitoring system.

[0132] S202. Based on historical meteorological data and a preset physical model, the theoretical power value and physical parameters are calculated.

[0133] Alternatively, one possible implementation for calculating the theoretical power value and physical parameters is as follows:

[0134] S2021. First, establish the preset physical model corresponding to the photovoltaic array. The preset physical model is shown in Formula 8:

[0135]

[0136] Formula 8 refers to the single-diode model equation of a photovoltaic cell, describing the relationship between the output current I and the output voltage V of the photovoltaic cell. In the formula, I refers to the output current of the photovoltaic cell, and V refers to the output voltage of the photovoltaic cell;ph This refers to the photocurrent, which is directly proportional to the light intensity; I0 refers to the reverse saturation current of the photovoltaic cell; R s Rsh refers to the series resistance; Rsh refers to the parallel resistance, i.e., the bypass resistance; n refers to the diode's ideality factor, typically between 1 and 2; V t This refers to thermal voltage.

[0137] S2022, Calculate effective irradiance based on historical meteorological data.

[0138] In this step, the effective irradiance is calculated as shown in Formula 9:

[0139]

[0140] Among them, G ' T refers to the effective irradiance, T refers to the ambient temperature in historical meteorological data, and G refers to the total irradiance of the photovoltaic array horizontal plane.

[0141] S2023. The theoretical power value corresponding to historical meteorological data is calculated based on the effective irradiance.

[0142] In this step, the theoretical power value is calculated as shown in Formula 10:

[0143]

[0144] Among them, P max This refers to the theoretical power value, also known as the maximum theoretical photovoltaic power generation, G. ' η refers to effective irradiance, η refers to photoelectric conversion efficiency, A is the total area of ​​the photovoltaic array, and T refers to the ambient temperature in historical meteorological data.

[0145] S2024. By combining Formula 8 and Formula 10, Formula 8 can be solved using the calculated maximum theoretical photovoltaic power generation, thereby obtaining the open-circuit voltage, short-circuit current, and fill factor corresponding to the maximum theoretical photovoltaic power generation. The open-circuit voltage, short-circuit current, and fill factor are then defined as physical parameters.

[0146] S203. Based on the theoretical power value, perform data processing and feature extraction on the historical power data to obtain sample training data.

[0147] Alternatively, one possible way to obtain sample training data is as follows:

[0148] S2031. Based on the theoretical power value, outlier filtering is performed on the historical power data to obtain historical power data with outliers removed.

[0149] In this step, the power range used for outlier screening is shown in Formula 11:

[0150]

[0151] Among them, P valid This refers to the effective power range. This refers to the photoelectric conversion efficiency; A refers to the total area of ​​the photovoltaic array, and G refers to the total area of ​​the photovoltaic array. max This refers to the historical maximum irradiance or the theoretical maximum irradiance; 1.2 refers to a safety factor, which means that the actual power is allowed to slightly exceed the theoretical maximum value.

[0152] Historical kilometer data that falls within this range are identified as normal values, while those that do not are identified as outliers, thus enabling the screening and removal of outliers.

[0153] S2032. For historical power data with outliers removed, fill in missing values ​​and normalize the data to obtain standard power data.

[0154] In this step, the missing values ​​are filled in as shown in Formula 12:

[0155]

[0156] Among them, P fill This refers to the power value after filling; P neighbor This refers to the power generation of other power plants adjacent to the current photovoltaic power plant at the same time; P physical This refers to the theoretical power value calculated using a pre-defined physical model.

[0157] In this step, the normalization process can be performed as follows:

[0158] a1. Convert horizontal irradiance to inclined irradiance for more accurate power generation calculation; the conversion method is based on the installation parameters of the photovoltaic array corresponding to the current historical power data, converting horizontal irradiance to inclined irradiance, as shown in Formula 13:

[0159]

[0160] in, The angle of incidence of sunlight. The zenith angle of the sun. G is the tilt angle of the photovoltaic panel. h This refers to horizontal irradiance, G diffuse This refers to the diffuse irradiance on a horizontal surface.

[0161] a2. The purpose of calculating the impact of module temperature on power generation efficiency is to correct the measured power based on the actual temperature of the solar cells, converting it to the power under standard test conditions; as shown in Formula 14:

[0162]

[0163] Among them, P corrected This refers to the corrected power; P measured This refers to the actual measured power in historical power data; T cell This refers to the actual temperature of the solar cells, T. STC This refers to the battery temperature under standard test conditions.

[0164] a3. Eliminating capacity differences between different power plants aims to divide the corrected power by the rated power to obtain the normalized power, as shown in Formula 15:

[0165]

[0166] Among them, P norm This refers to the normalized power; P corrected This refers to the corrected power, P. rated This refers to the rated power of a photovoltaic power station.

[0167] S2033. Multi-scale periodic feature extraction is performed based on standard power data to obtain sample training data.

[0168] In this step, the process of multi-scale periodic feature extraction is as follows:

[0169] b1. The transient fluctuation characteristics of photovoltaic power are extracted using short-time Fourier transform, as shown in Equation 16:

[0170]

[0171] in, Let P(t) be the time-frequency energy density, and P(t) be the photovoltaic power time-series signal. is the Hanning window function; f is the frequency component, i.e., the time-frequency energy density.

[0172] b2. The daily cycle phase difference is used to reflect the characteristics of the daily cycle, as shown in Formula 17:

[0173]

[0174] in, For the optimal phase difference, This is the phase offset. The ideal daily cycle template refers to the ideal daily cycle power curve. The purpose of this step is to reflect the phase deviation between the actual daily cycle and the ideal daily cycle, and to determine the phase difference of the daily cycle as a characteristic of the daily cycle.

[0175] b3. Extract the long-term trend using a 30-day moving average, as shown in Formula 18:

[0176]

[0177] Among them, P trend P(tk) represents the seasonal trend term, and P(tk) represents the power on the k-th historical day.

[0178] b4. Combine the above process to extract periodic features across the entire scale, and integrate the periodic features from three different time scales into sample training data.

[0179] S204. Based on the theoretical power value, physical parameters, and sample training data, the initial mathematical model is iteratively distilled and trained to obtain the target mathematical model.

[0180] Alternatively, one possible implementation of obtaining the target mathematical model through iterative distillation training is as follows:

[0181] S2041. In each iteration of distillation training, based on the periodic encoding layer of the initial mathematical model, each multi-scale periodic feature data in the sample training data is transformed into the initial hidden state.

[0182] In this step, the multi-scale periodic feature data can be transformed into an initial hidden state based on the periodic coding layer, as shown in Equation 19:

[0183]

[0184] MLP stands for Multilayer Perceptron. Indicates feature concatenation, f short The dominant frequency of short-term fluctuations is derived from short-time Fourier transform analysis. P represents the daily periodic phase difference. trend As a seasonal trend component, the weights of each cyclical component are dynamically adjusted through a cyclical attention mechanism.

[0185] The purpose of this step is to transform multi-scale periodic features into an initial state that the neural network can utilize efficiently, guiding the model to focus on key time patterns.

[0186] S2042. Based on the TimesNet core layer in the initial mathematical model, time-frequency transformation is performed on the initial hidden state and multi-scale periodic features to obtain the time-series feature vector.

[0187] In this step, the time-frequency transformation process is as follows: based on the initial hidden state and multi-scale periodic features, a time-frequency transformation is performed to convert the time series from the time domain to the frequency domain; the frequency domain representation is reshaped into a 2D structure, and then a 2D convolutional neural network is applied for feature extraction; the processed frequency domain data is then converted back to the time domain to obtain the time series feature vector. Furthermore, the TimesNet core layer uses the initial hidden state h0 as the initial state of the recurrent neural network to guide the time series processing; and dynamically adjusts the weights of different periodic components through a periodic attention mechanism.

[0188] S2043. Calculate the original predicted output value corresponding to the temporal feature vector based on the fully connected layer in the initial mathematical model.

[0189] The purpose of this step is to pass the temporal feature vector through a fully connected layer to obtain the original prediction output value. The method is to use the linear transformation of the fully connected layer to convert the temporal feature vector into a scalar, thereby obtaining the original prediction output value.

[0190] S2044. Constrain the original predicted output value based on the physical parameters to obtain the target predicted output value.

[0191] Alternatively, the constraint handling can be performed as shown in Equations 20 and 21:

[0192]

[0193] in, The final predicted power corresponds to the target predicted output value. , , These are open-circuit voltage, short-circuit current, and All of these are physical parameters calculated from a preset physical model; WX+b refers to the original predicted output value. "Sigmoid" refers to the sigmoid function; "tanh" refers to the hyperbolic tangent function.

[0194] S2045. Distillation training is performed based on the theoretical power value and the target predicted output value to obtain the target mathematical model.

[0195] Alternatively, one possible implementation of distillation training to obtain the target mathematical model is as follows:

[0196] A1. Based on the theoretical power value and the target predicted output value, perform feature decomposition to obtain the first short-term component, the first daily cycle component, and the first long-term component corresponding to the theoretical power value, and the second short-term component, the second daily cycle component, and the second long-term component corresponding to the target predicted output value.

[0197] In this step, the short-term component refers to the high-frequency fluctuations extracted using a high-pass filter; the daily cycle component refers to the daily cycle-related fluctuations extracted using a band-pass filter; and the long-term component refers to the long-term trend extracted using a low-pass filter.

[0198] For example, one possible implementation of feature decomposition is as follows:

[0199] d1. Filter the theoretical power value and the target predicted output value to obtain their respective components.

[0200] d2. For short-term components, use a high-pass filter with a cutoff frequency of 1 / 6 hour.

[0201] d3. For the daily periodic component, use a bandpass filter with a low-frequency cutoff of 1 / 32 hour and a high-frequency cutoff of 1 / 6 hour.

[0202] d4. For long-term components, use a low-pass filter or a low-pass filter with a cutoff frequency of 1 / 48 of an hour.

[0203] A2. Based on the first short-term component, the first daily cycle component, the first long-term component, the second short-term component, the second daily cycle component, and the second long-term component, the distillation loss is calculated.

[0204] In this step, the distillation loss is calculated as shown in Formula 22:

[0205]

[0206] Among them, L KD This refers to distillation loss; F k F refers to the function that extracts the k-th time-scale component from the power sequence. s F represents the short-term fluctuation component. d F represents the daily periodic component. l Indicates a long-term trend; P tea This refers to the power sequence output by the preset physical model; P stu This refers to the power sequence output by the mathematical model; This refers to the distillation intensity.

[0207] Distillation strength It is adjusted in real time based on meteorological conditions, as shown in Formula 23:

[0208]

[0209] Where C represents the confidence level of meteorological classification, sunny days emphasize daily cycle alignment, cloudy days strengthen short-term fluctuation matching, and overcast and rainy days enhance the consistency of trend terms; W k For learnable parameter vectors, This refers to the distillation intensity.

[0210] A3. The prediction loss is calculated based on the target prediction output value and the original output value corresponding to the sample training data.

[0211] In this step, the prediction loss refers to the difference between the predicted output value of the mathematical model and the original output value in the sample training data. Various types of loss calculation functions can be used to calculate the prediction loss corresponding to this difference.

[0212] A4. The total loss is obtained by weighting the physical regularization input, prediction loss and distillation loss corresponding to the preset physical model.

[0213] In this step, the physical equations of the preset physical model are injected as regularization terms, as shown in Equation 24:

[0214]

[0215] This formula is used to calculate the predicted power of the mathematical model against the parameters. The norm difference between gradients in the pre-defined physical model. phy This refers to the physical regularization term; This refers to the gradient of the power output by the mathematical model with respect to the model parameters; J tea This refers to the gradient of the power output by the preset physical model with respect to the input meteorological data, i.e., the Jacobian matrix of the preset physical model; IIII F It refers to the square root of the sum of the squares of all elements.

[0216] In this step, the total loss is calculated as shown in Formula 25:

[0217]

[0218] in, This refers to the weighting coefficient of the distillation loss, which decays with each training round; This refers to the fixed weight of the physical regularization term; L pred This refers to predicted loss; L phy This refers to the physical regularization term; L KD This refers to distillation loss; L refers to total loss.

[0219] A5. When the total loss is less than the preset threshold, the mathematical model in the current training phase is determined as the target mathematical model.

[0220] In this step, when the total loss is lower than a preset threshold, iterative training can be stopped and the mathematical model of the current training stage can be determined as the target mathematical model; iterative training can also be stopped and the target mathematical model determined when the number of iterations reaches the upper limit.

[0221] Figure 3 A flowchart illustrating the photovoltaic power generation prediction method provided in this application. Figure 3 ,like Figure 3 As shown, the method includes:

[0222] B1. Construct a physical model and calculate theoretical power based on the physical characteristics of photovoltaics.

[0223] The specific steps include: collecting meteorological data; constructing a physical model based on the physical characteristics of the photovoltaic array; calculating physical parameters based on the physical model; and calculating the effective irradiance and theoretical power value.

[0224] B2. Data preprocessing and periodic feature enhancement.

[0225] The specific steps include: acquiring raw power data; performing dynamic threshold anomaly detection based on theoretical power values; calculating spatiotemporal mixing differences for missing values; correcting power based on temperature; and extracting instantaneous fluctuation features, daily cycle features, and long-term trends.

[0226] B3. Optimize the TimeNet mathematical model based on the fusion of physical constraints and multi-scale time-frequency features.

[0227] The specific steps include: adding a periodic encoding layer; performing feature state transformation based on the periodic encoding layer to obtain the initial hidden state corresponding to the multi-out-degree periodic features; dynamically adjusting the weights of the periodic components; limiting the prediction power to a physically feasible range based on physical constraints; and limiting the prediction power to obtain the prediction power output.

[0228] B4. Periodic knowledge distillation of multi-scale video features.

[0229] The specific steps include: constructing a distillation loss function; performing feature decomposition and matching on the predicted output of the mathematical model and the theoretical power of the physical model to obtain matching pairs; calculating the distillation loss based on the distillation loss function and the matching pairs; adjusting the distillation intensity in real time; injecting the physical equations of the physical model and calculating the physical regularization term; and calculating the total loss based on the distillation loss, the physical regularization term, and the predicted loss of the mathematical model.

[0230] B5. Hybrid prediction of photovoltaic power driven by physical and data dual-drive based on dynamic weight optimization.

[0231] The specific steps include: dual-channel parallel prediction; prediction based on physical channels and data-driven channels, and calculation of corrected power; calculation of dynamic weights, and weighted calculation based on dynamic weights to obtain mixed power; and output of mixed power.

[0232] Figure 4 This is a schematic diagram of the photovoltaic power generation prediction device provided in this application, as shown below. Figure 4As shown, the photovoltaic power generation prediction device provided in this embodiment includes:

[0233] The acquisition module 401 is used to acquire real-time meteorological data and the corresponding power time-series data.

[0234] The first processing module 402 is used to calculate the theoretical power generation corresponding to real-time meteorological data based on a preset physical model.

[0235] The second processing module 403 is used to calculate the predicted power generation corresponding to the power time series data based on a pre-trained mathematical model.

[0236] The third processing module 404 is used to calculate the corrected power based on the theoretical power generation and the predicted power generation.

[0237] The fourth processing module 405 is used to calculate dynamic weights based on real-time meteorological data and corrected power.

[0238] The fifth processing module 406 is used to perform weighted calculations based on dynamic weights, corrected power, and theoretical power generation to obtain the target predicted power corresponding to the real-time meteorological data.

[0239] In one possible implementation, the fourth processing module 405 is further used for:

[0240] Based on real-time meteorological data and corrected power, the confidence scores of the physical model and the mathematical model are calculated.

[0241] Based on preset hyperparameters, the confidence scores of the physical model and the mathematical model are weighted and normalized to obtain dynamic weights.

[0242] In one possible implementation, the device further includes a sixth processing module 407, for:

[0243] Acquire historical meteorological data and historical power data.

[0244] Based on historical meteorological data and a pre-set physical model, the theoretical power value and physical parameters are calculated.

[0245] Based on theoretical power values, historical power data are processed and features are extracted to obtain sample training data.

[0246] The initial mathematical model is iteratively distilled and trained based on theoretical power values, physical parameters, and sample training data to obtain the target mathematical model.

[0247] In one possible implementation, the sixth processing module 407 is further configured to:

[0248] Outlier filtering is performed on historical power data based on theoretical power values ​​to obtain historical power data with outliers removed.

[0249] The historical power data, after removing outliers, is filled with missing values ​​and normalized to obtain standard power data.

[0250] Multi-scale periodic feature extraction is performed based on standard power data to obtain sample training data.

[0251] In one possible implementation, the sixth processing module 407 is further configured to:

[0252] In each iteration of distillation training, the periodic encoding layer based on the initial mathematical model transforms each multi-scale periodic feature data in the sample training data into an initial hidden state.

[0253] Based on the TimesNet core layer in the initial mathematical model, time-frequency transformation is performed on the initial hidden state and multi-scale periodic features to obtain the time-series feature vector.

[0254] The original predicted output value corresponding to the temporal feature vector is calculated based on the fully connected layer in the initial mathematical model.

[0255] The original predicted output value is constrained based on physical parameters to obtain the target predicted output value.

[0256] Distillation training is performed based on the theoretical power value and the target predicted output value to obtain the target mathematical model.

[0257] In one possible implementation, the sixth processing module 407 is further configured to:

[0258] Based on the theoretical power value and the target predicted output value, feature decomposition is performed to obtain the first short-term component, the first daily cycle component, and the first long-term component corresponding to the theoretical power value, and the second short-term component, the second daily cycle component, and the second long-term component corresponding to the target predicted output value.

[0259] The distillation loss is calculated based on the first short-term component, the first daily periodic component, the first long-term component, the second short-term component, the second daily periodic component, and the second long-term component.

[0260] The prediction loss is calculated based on the target prediction output value and the original output value corresponding to the sample training data.

[0261] The total loss is obtained by weighting the physical regularization input, prediction loss, and distillation loss corresponding to the preset physical model.

[0262] When the total loss is less than a preset threshold, the mathematical model in the current training phase is determined as the target mathematical model.

[0263] The apparatus provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0264] Figure 5 A schematic diagram of the structure of the electronic device provided in this application. Figure 5 As shown, the electronic device provided in this embodiment includes at least one processor 501 and a memory 502. Optionally, the device further includes a communication component 503. The processor 501, memory 502, and communication component 503 are connected via a bus 504.

[0265] In the specific implementation process, at least one processor 501 executes computer execution instructions stored in memory 502, causing at least one processor 501 to execute the above-mentioned photovoltaic power generation prediction method or method.

[0266] The specific implementation process of processor 501 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0267] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0268] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0269] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0270] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0271] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0272] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0273] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0274] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0275] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0276] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0277] If a function 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 invention, or the part that contributes to the prior art, or a 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 of the various embodiments of this invention. 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.

[0278] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0279] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for predicting photovoltaic power generation, characterized in that, include: Acquire real-time meteorological data and the corresponding power time-series data; The theoretical power generation corresponding to the real-time meteorological data is calculated based on a preset physical model. The predicted power generation corresponding to the power time series data is calculated based on a pre-trained mathematical model; The corrected power is calculated based on the theoretical power generation and the predicted power generation. Based on the real-time meteorological data and the corrected power, the dynamic weight is calculated; The target predicted power corresponding to the real-time meteorological data is obtained by weighting the dynamic weight, the corrected power, and the theoretical power generation.

2. The method according to claim 1, characterized in that, The calculation of dynamic weights based on the real-time meteorological data and the corrected power includes: Based on the real-time meteorological data and the corrected power, the confidence scores of the physical model and the mathematical model are calculated. Based on preset hyperparameters, the confidence scores of the physical model and the mathematical model are weighted and normalized to obtain the dynamic weights.

3. The method according to claim 1, characterized in that, Before acquiring real-time meteorological data and the corresponding power time-series data, the method further includes: Acquire historical meteorological data and historical power data; Based on the historical meteorological data and the preset physical model, the theoretical power value and physical parameters are calculated. Based on the theoretical power value, the historical power data is processed and features are extracted to obtain sample training data; Based on the theoretical power value, the physical parameters, and the sample training data, the initial mathematical model is iteratively distilled and trained to obtain the target mathematical model.

4. The method according to claim 3, characterized in that, The process of data processing and feature extraction on the historical power data based on the theoretical power value to obtain sample training data includes: Based on the theoretical power value, outlier filtering is performed on the historical power data to obtain historical power data with outliers removed. The historical power data, after removing outliers, is filled with missing values ​​and normalized to obtain standard power data. Multi-scale periodic feature extraction is performed based on the standard power data to obtain sample training data.

5. The method according to claim 3, characterized in that, The step of iteratively distilling and training the initial mathematical model based on the theoretical power value, the physical parameters, and the sample training data to obtain the target mathematical model includes: In each iteration of distillation training, based on the periodic encoding layer of the initial mathematical model, each multi-scale periodic feature data in the sample training data is transformed into an initial hidden state. Based on the TimesNet core layer in the initial mathematical model, the initial hidden state and the multi-scale periodic features are subjected to time-frequency transformation to obtain a time-series feature vector. The original predicted output value corresponding to the temporal feature vector is calculated based on the fully connected layer in the initial mathematical model. The original predicted output value is constrained based on the physical parameters to obtain the target predicted output value. Distillation training is performed based on the theoretical power value and the target predicted output value to obtain the target mathematical model.

6. The method according to claim 3, characterized in that, The process of distilling and training based on the theoretical power value and the target predicted output value to obtain the target mathematical model includes: Based on the theoretical power value and the target predicted output value, feature decomposition is performed to obtain the first short-term component, the first daily cycle component, and the first long-term component corresponding to the theoretical power value, and the second short-term component, the second daily cycle component, and the second long-term component corresponding to the target predicted output value. Based on the first short-term component, the first daily periodic component, the first long-term component, the two short-term components, the second daily periodic component, and the second long-term component, the distillation loss is calculated. The prediction loss is calculated based on the target prediction output value and the original output value corresponding to the sample training data. The total loss is obtained by weighting the physical regularization input corresponding to the preset physical model, the prediction loss, and the distillation loss. When the total loss is less than a preset threshold, the mathematical model in the current training phase is determined as the target mathematical model.

7. A photovoltaic power generation prediction device, characterized in that, include: The acquisition module is used to acquire real-time meteorological data and the corresponding power time-series data. The first processing module is used to calculate the theoretical power generation corresponding to the real-time meteorological data based on a preset physical model. The second processing module is used to calculate the predicted power generation corresponding to the power time series data based on a pre-trained mathematical model. The third processing module is used to calculate the corrected power based on the theoretical power generation and the predicted power generation. The fourth processing module is used to calculate dynamic weights based on the real-time meteorological data and the corrected power. The fifth processing module is used to perform weighted calculations based on the dynamic weights, the corrected power, and the theoretical power generation to obtain the target predicted power corresponding to the real-time meteorological data.

8. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-6.

10. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method described in any one of claims 1-6.

Images