Physical uncertainty guided hybrid photovoltaic power interval prediction method and system

Abstract

The invention provides a hybrid photovoltaic power interval prediction method and system guided by physical uncertainty, and relates to the field of power system operation and maintenance and energy storage deployment, and the method comprises the following steps: obtaining the historical actual measurement data of a photovoltaic power station, and constructing a multi-dimensional time series data set for hybrid modeling; a photovoltaic physical model chain is constructed, the sun geometric position, the slope irradiance, the battery temperature and the photoelectric conversion efficiency are calculated in sequence, and meteorological data are mapped into deterministic reference power. A dynamic uncertainty quantification mechanism is introduced, a machine learning model is utilized to predict the variance of input meteorological factors, and the theoretical standard deviation of the physical model is calculated based on the error propagation law in combination with the instantaneous sensitivity of the physical model to each input; and constructing a mixed feature vector, and learning residual distribution under different quantiles. And synthesizing the synthetic physical reference power and the predicted residual quantile to generate a final photovoltaic power probability prediction interval, and outputting a high-precision prediction result which is guided by physical characteristics and corrected by deep learning.

Description

Technical Field

[0001] This invention relates to the field of power system maintenance and energy storage and dispatch. Background Technology

[0002] As the global energy structure transitions towards low-carbon and clean energy, the installed capacity of new energy sources, represented by photovoltaics, continues to rise in the power system. However, photovoltaic power generation is highly volatile, intermittent, and random, with its output significantly affected by meteorological factors such as cloud movement, aerosol variations, temperature, and wind speed. Large-scale photovoltaic grid connection poses severe challenges to grid power balance, spinning reserve arrangements, and electricity market transactions. To address this challenge, high-precision, quantifiable uncertainty-based photovoltaic power probability prediction technology has become a key supporting technology for the safe and stable operation of new power systems.

[0003] Currently, photovoltaic (PV) power prediction technologies are mainly divided into physical model methods and data-driven methods, but each has its own insurmountable limitations. Physical model methods, based on atmospheric physics and photoelectric conversion principles, establish a complete physical model chain from numerical weather prediction (NWP) to PV power. Their advantages include clear physical meaning, the ability to cold-start without extensive historical data, and strong extrapolation capabilities under extreme weather conditions. However, physical models are "white-box" models, relying on precise power plant parameters. In actual operation, many structural errors are difficult to model accurately, such as complex shading, component dust accumulation, spectral mismatch, and inverter nonlinear peak clipping. These unmodeled physical processes lead to systematic biases in prediction results, which are difficult to eliminate through simple parameter adjustments. Data-driven methods utilize deep learning algorithms to directly establish a nonlinear mapping between meteorological and power data. While they can capture complex patterns in the data through powerful fitting capabilities, they are "black-box" models, lacking physical constraints. In extreme weather scenarios not covered by training data, they are prone to predictions that violate physical principles, and the sources of error are difficult to explain.

[0004] To combine the advantages of both approaches, the industry has begun exploring a hybrid "physics + data" model. Existing hybrid methods primarily focus on input-side calibration, using machine learning to correct NWP data before inputting it into the physical model. However, this method of merely calibrating the input meteorological data cannot address the structural defects of the physical model itself; the output residuals of the physical model often contain complex nonlinear laws. Furthermore, existing residual correction models typically treat the physical model's output as merely a general feature, or directly ignore the physical model's "confidence" in the current operating conditions. In fact, the error propagation mechanism of the physical model differs under different weather conditions. For example, on sunny days, the prediction variance of the physical model is usually small, while on cloudy days, due to its sensitivity to cloud changes, the uncertainty of the prediction results is drastically amplified. When predicting residuals, existing general-purpose deep learning models, lacking prior knowledge from the physics domain, struggle to distinguish between random noise and physical model failure as the causes of bias, resulting in prediction intervals that are often too wide or too narrow, lacking specificity.

[0005] To address the above issues, there is an urgent need for a method that can deeply integrate physical mechanisms with deep learning. Summary of the Invention

[0006] To overcome the problem of unreasonable prediction ranges in existing prediction methods, this invention provides a hybrid power range prediction method and system guided by physical uncertainty.

[0007] The technical solution adopted by this invention to achieve the above objectives is: a hybrid photovoltaic power range prediction method guided by physical uncertainty, comprising the following steps:

[0008] Acquire meteorological driving data from numerical weather prediction (NWP) and corresponding historical measured data from photovoltaic power plants; perform time zone alignment and resampling operations to address the spatiotemporal resolution differences between data sources. Construct time-series samples using a sliding window mechanism: set the historical backtracking window length to... The meteorological and power data from consecutive time points are sliced ​​to construct a shape of... A high-dimensional temporal tensor is used. Simultaneously, to ensure training stability for deep neural networks, Z-Score normalization is performed on all features, mapping data with vastly different physical dimensions to a standard normal distribution, forming a normalized temporal dataset to drive subsequent deep learning models.

[0009] A photovoltaic physical model chain is constructed using a normalized time-series dataset to sequentially calculate the solar geometric position, slope irradiance, cell temperature, and photoelectric conversion efficiency, mapping the NWP meteorological data into a deterministic reference power; this reference power represents the predicted value under an ideal physical description.

[0010] By introducing a dynamic uncertainty quantification mechanism using the theoretical standard deviation sequence of the above physical model, a machine learning model is used to predict the variance of the input meteorological factors, and the instantaneous sensitivity (gradient) of the physical model to each input is combined with the error propagation law to calculate the theoretical standard deviation of the physical model. This index quantifies the sensitivity and confidence of the physical model to the current meteorological conditions.

[0011] A physical-guided residual correction dataset is constructed, and the difference between the measured power and the reference power is calculated to obtain the residual sequence. A hybrid feature vector is constructed, which is forced to include the obtained reference power, the theoretical standard deviation obtained in step 3, and the original NWP meteorological features. A quantile recurrent neural network is used with the hybrid features as input and the residual sequence as the target to learn the residual distribution under different quantiles.

[0012] The synthesized physical baseline power and the predicted residual quantiles are used to generate the final photovoltaic power probability prediction interval; through non-negativity truncation and evaluation index verification, the prediction results guided by physical features and corrected by deep learning are output.

[0013] Preferably, in the data preprocessing, to address the issue of inconsistent time zones between data from different sources (e.g., NWP data is usually in UTC time, while measured data is in local time), the local time zone (UTC+8) where the photovoltaic power station's latitude and longitude are located is read, and the timestamp of the NWP data is added with the corresponding time difference offset to ensure that it is strictly aligned with the measured records on the time axis.

[0014] For the shortwave radiation and cumulative surface thermal radiation data output by NWP, the data are divided by the time interval according to the time step of NWP to convert them into average power flux, so that their physical dimensions are consistent with the input requirements of the photovoltaic physical model.

[0015] Furthermore, abnormal data is eliminated based on physical extremes and statistical regularities. The elimination rules include: (1) eliminating nighttime data (based on a solar zenith angle greater than 85°); (2) eliminating data that violates physical limits (such as irradiance exceeding the solar constant or power exceeding the installed capacity); and (3) eliminating "dead values" caused by equipment failure (values ​​that remain unchanged for a long period of time). The rules are based on the physical principles of photovoltaic power generation and the nominal parameters of the equipment to ensure the authenticity and validity of the modeling data.

[0016] The aforementioned time-series datasets include meteorological driving data based on numerical weather forecasting and corresponding historical measured data from photovoltaic power plants.

[0017] Preferably, when constructing the physical model chain, the total irradiance of the inclined plane is calculated. :

[0018] ;

[0019] In the formula, for Total irradiance on the inclined surface of the photovoltaic array at any given time; The component of direct irradiation on the inclined plane; This represents the irradiance component scattered by the inclined plane. This represents the irradiance component reflected from the sloping ground.

[0020] Calculate the temperature of photovoltaic cells :

[0021] ;

[0022] In the formula, The baseline DC power calculated for the physical model; The rated capacity of the photovoltaic array under standard test conditions; The power temperature coefficient; This is the overall system efficiency coefficient, which includes inverter efficiency and line loss.

[0023] Preferably, the dynamic uncertainty quantification mechanism and error propagation law can be decomposed into the following three specific mathematical calculation steps:

[0024] Preferably, the uncertainty of meteorological factors is modeled using machine learning models (such as Gaussian process regression or probability-based neural networks) for key meteorological factors of NWP (total irradiance). ,temperature Modeling is performed.

[0025] Using gradient boosting trees (LightGBM) based on weather state features Predicting various meteorological input variables conditional variance of :

[0026] ;

[0027] In the formula, For the first Each meteorological input variable (such as DNI, wind speed) in The variance of the prediction at any given time; For the trained gradient boosting tree model; The input is the state feature vector; the model outputs not only predicted values ​​of meteorological factors. It also outputs the variance of the predicted value. .

[0028] Preferably, the partial derivatives (gradients) of the physical model output with respect to each input variable are calculated using the numerical difference method:

[0029] ;

[0030] In the formula, for Power at time t = 1 The physical sensitivity gradient of each variable; It represents a tiny numerical perturbation.

[0031] Preferably, the total variance of the synthetic physics derivation The theoretical standard deviation is obtained by taking the square root:

[0032] ;

[0033] In the formula, The theoretical standard deviation derived from the physical model will be passed as a key feature to subsequent steps.

[0034] Preferably, constructing the residual correction dataset specifically involves defining the target residual. :

[0035] ;

[0036] In the formula, For true residuals; This is the measured photovoltaic power.

[0037] Constructing a hybrid feature vector The data is then input into a quantile LSTM network for training, with the objective function being a comprehensive bouncing loss function. :

[0038] ;

[0039] In the formula, Set the target quantiles (e.g., 0.025, 0.5, 0.975); The model predicts the first The residual value of the quantile.

[0040] Preferably, the synthetic prediction calculation formula is as follows:

[0041] ;

[0042] In the formula, The final confidence level is The power prediction boundary value.

[0043] Preferably, a hybrid feature vector is constructed. The process involves a combination of feature engineering and physical embedding, and the specific process is as follows:

[0044] Preferably, the selection and definition of feature dimensions: hybrid feature vector At time step It consists of the following three heterogeneous information components, aiming to complement the strengths and weaknesses of physical models and data-driven models:

[0045] The first part is derived from the deterministic reference power calculated by the photovoltaic physical model chain as a physical reference feature. This provides the main trend and physical lower bound of power variation; it tells the neural network "how much power should be under ideal physical conditions".

[0046] The second part is derived from the theoretical standard deviation calculated based on the law of error propagation, which serves as the physical confidence characteristic. The value is used as attention weights or gating signals; the larger the value, the less reliable the physical model's prediction of the current moment (such as when the cloud layer changes rapidly), suggesting that the neural network should rely more on meteorological data for correction; the smaller the value, the more reliable the physical model is, and the network should retain the physical baseline value.

[0047] Part Three: Characteristics of the Primitive Meteorological Environment It provides environmental context information to help neural networks capture nonlinear weather patterns that physical models fail to model.

[0048] Preferably, the above three parts of data are concatenated along the feature dimension; assuming time... meteorological characteristics Dimensions The construction process is shown in the following formula:

[0049] ;

[0050] concatenated feature vectors The total dimension is .

[0051] Preferably, in order to eliminate the training instability caused by different physical dimensions, standardization is performed: normalization: normalization is applied to the concatenated vectors. Each feature dimension in the dataset is Z-score standardized (mean subtracted, standard deviation divided) to make its distribution approximate. To capture time dependencies, it's not necessary to input only a single moment. Instead, it constructs a time window sequence. This is used as an input sample for the neural network.

[0052] A hybrid photovoltaic power range prediction system guided by physical uncertainties includes:

[0053] The data preprocessing module is used to acquire meteorological driving data from numerical weather forecasts and corresponding historical measured data from photovoltaic power plants; to uniformly correct the time zones of data from different sources; to convert cumulative quantities such as shortwave radiation into instantaneous quantities and align them with the measured records along the time axis; and to remove outlier data.

[0054] The power calculation module, connected to the data preprocessing module, is used to calculate the solar geometric position, slope irradiance, battery temperature and photoelectric conversion efficiency, and to map the above numerical weather forecast meteorological data into a deterministic reference power.

[0055] The machine learning module, connected to the power calculation module, calculates the difference between the measured power and the reference power to obtain the residual sequence; constructs a mixed feature vector and the original NWP meteorological features; and uses a quantile recurrent neural network with the mixed features as input and the residual sequence as the target to learn the residual distribution under different quantiles.

[0056] The output module is connected to the machine learning module. It combines the residual distribution characteristics obtained by machine learning and synthesizes the physical baseline power and the predicted residual quantile to generate the final photovoltaic power probability prediction interval. It outputs the prediction results guided by physical features and corrected by deep learning.

[0057] The power supply module is electrically connected to the data preprocessing module, power calculation module, machine learning module, and output module, respectively, and provides power to the data preprocessing module, power calculation module, machine learning module, and output module.

[0058] A computer-readable storage medium for an uncertainty-guided hybrid photovoltaic power range prediction system stores readable computer instructions for performing the steps of the power range prediction method.

[0059] This invention constructs a hybrid framework of "physical derivation + residual correction," utilizing a chain of physical models to calculate baseline power and theoretical variance. These intermediate physical quantities are then input as strong prior features into a quantile recurrent neural network, specifically learning the nonlinear residual distribution that the physical model fails to capture. These probabilistic prediction results, incorporating physical information, can significantly improve prediction accuracy and inter-regional reliability under complex weather conditions, providing more valuable risk decision-making basis for grid dispatching, thereby enhancing the grid's capacity to absorb renewable energy.

[0060] By directly modeling the output residuals, this method effectively captures and corrects systematic biases in the physical model caused by inaccurately described physical processes such as shading, component aging, dust accumulation, or inverter nonlinear efficiency. This "physical feature-guided" mechanism enables the neural network to perceive the "confidence level" of the physical model—when the physical model is unreliable, the neural network automatically adjusts the residual prediction range, thereby achieving a complementary advantage between physical mechanisms and data-driven approaches. This solves the problem of weak generalization ability of traditional "black box" models under extreme weather conditions, retaining the high accuracy of the physical model under clear weather conditions while correcting nonlinear biases under complex weather conditions through deep learning. Attached Figure Description

[0061] Figure 1 This is a flowchart of an embodiment of the present invention.

[0062] Figure 2 This is a system principle block diagram of an embodiment of the present invention. Detailed Implementation

[0063] The physical uncertainty-guided hybrid photovoltaic power range prediction method of the present invention includes the following steps:

[0064] Step 1: Obtain numerical weather prediction (NWP) meteorological driving data and corresponding measured photovoltaic power and irradiance data, preprocess the data, and construct a multi-source fusion dataset.

[0065] This process first requires addressing the time reference discrepancies between different data sources. The Coordinated Universal Time (UTC) of the NWP data is converted to the local time consistent with the measured data, ensuring strict alignment in spatiotemporal reference. For the common shortwave radiation accumulation in NWP data, averaging is performed based on the time step of its data release, converting it into an average irradiance with physical instantaneous significance. Then, interpolation algorithms are used to resample data at different time resolutions to maintain consistency with the temporal granularity of the measured data.

[0066] Based on this, strict quality control is carried out on the measured data. Statistical methods are used to identify and remove continuous dead value data and non-zero abnormal drift data that appear at night, so as to ensure the purity of the training data.

[0067] To enhance the model's ability to perceive different weather conditions, the theoretical clear sky irradiance is calculated using the clear sky model, and a clear sky index is constructed by comparing it with the actual irradiance. This index is used to classify historical data into different weather types, providing state labels for subsequent scenario-based modeling.

[0068] Step 2: Construct a simulation model chain encompassing the entire physical process of photovoltaic modules, mapping meteorological data into deterministic reference power. The core of this step lies in simulating the photon-to-electron conversion process using physical equations.

[0069] Step 2-1: Calculate the geometric position of the sun, including the altitude angle and azimuth angle, based on geographical latitude and longitude and time information. Then, use the irradiance separation model to decompose the total irradiance on the horizontal plane into direct component, scattered component and ground reflection component. Combined with the installation tilt angle and azimuth angle of the photovoltaic array, calculate the effective total irradiance on the inclined surface of the photovoltaic module.

[0070] Step 2-2 introduces a thermodynamic model that takes into account environmental heat exchange, and uses ambient temperature and wind speed data to estimate the actual operating temperature of the photovoltaic cell.

[0071] Steps 2-3 involve combining the rated capacity of the photovoltaic modules, the power temperature coefficient, and the overall system efficiency, inputting the slope irradiance and cell temperature into the photoelectric conversion model to calculate a deterministic reference power. This reference power represents the theoretical power generation under ideal physical conditions, assuming no structural biases, and will serve as the predictive framework for subsequent hybrid models.

[0072] Step 3: Introduce a dynamic uncertainty quantification mechanism to derive the theoretical standard deviation of the physical model, so as to quantify the physical model's "confidence" in the current meteorological conditions.

[0073] Step 3-1: It is necessary to establish an error prediction model that can sense weather conditions. For example, by using the gradient boosting tree algorithm, a mapping relationship between NWP meteorological characteristics and NWP prediction error can be established, so as to dynamically predict the variance of each meteorological input variable (such as irradiance, wind speed, etc.) at each moment.

[0074] Step 3-2: Using the idea of ​​numerical differentiation, apply a small perturbation to each input variable in the physical model chain, observe the change in output power, and thus calculate the instantaneous sensitivity of the physical model output to each input variable (i.e., physical gradient).

[0075] Step 3-3: Based on the error propagation law, the predicted variance of each input variable is weighted and synthesized with the corresponding physical sensitivity to calculate the total variance derived from the physical model and then take the square root to obtain the theoretical standard deviation. This indicator has a clear physical meaning: when the weather forecast is inaccurate and the physical model is very sensitive to the weather factor, the theoretical standard deviation will increase significantly, and vice versa.

[0076] Steps 3-4: Construct a residual correction dataset guided by physical features, and use a quantile recurrent neural network to learn the conditional probability distribution of the residuals.

[0077] Step 4: Calculate the difference between the measured power and the reference power obtained in Step 2 to obtain the target residual sequence. This sequence contains all nonlinear deviations (such as shadow occlusion, dust accumulation, etc.) that the physical model failed to capture.

[0078] Step 4-1: Construct a hybrid feature vector that not only includes the original meteorological and temporal features, but more importantly, forcibly incorporates the baseline power calculated in Step 2 and the theoretical standard deviation calculated in Step 3. These two physical features act as a "guide," informing the neural network physical model of the current prediction value and the degree of uncertainty the physical model perceives in that prediction.

[0079] Step 4-2: Establish a quantile long short-term memory network (Quantile LSTM), using the mixed feature sequence as input and the residual sequence as the target for training.

[0080] Step 4-3: In order to achieve interval prediction, the model is optimized using a comprehensive quantile loss function (such as bouncing loss), which forces the neural network to output multiple quantiles of the residual distribution (such as the lower and upper boundaries) at the same time, thereby accurately characterizing the error range of the physical model under different working conditions.

[0081] Step 5: Combine the physical reference power with the predicted residual quantiles to generate the final photovoltaic power probability prediction interval.

[0082] Step 5-1: Directly add the deterministic reference power calculated in Step 2 to the quantile residual values ​​output by the neural network in Step 4. For example, the lower bound of the prediction interval is formed by the reference power plus the predicted low quantile residual, and the upper bound is formed by the reference power plus the predicted high quantile residual. To conform to the physical characteristics of photovoltaic power generation, a non-negativity constraint is applied to the synthesized result, i.e., when the calculated result is less than zero, it is forcibly set to zero.

[0083] Step 5-2: The final output prediction interval retains the high accuracy of the physical model under clear weather conditions, and corrects the nonlinear structural bias under complex weather conditions through deep learning. This achieves a deep integration of physical mechanisms and data-driven approaches, providing highly reliable probabilistic information for power grid dispatch.

[0084] This invention has been described through embodiments. Those skilled in the art will understand that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of this invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, this invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of this invention.

Claims

1. A hybrid photovoltaic power range prediction method guided by physical uncertainty, characterized in that, include: Acquire meteorological driving data from numerical weather forecasts and corresponding historical measured data from photovoltaic power plants to construct a multidimensional time-series dataset for hybrid modeling. A photovoltaic physical model chain is constructed to calculate the solar geometric position, slope irradiance, cell temperature and photoelectric conversion efficiency in sequence, and the numerical weather forecast meteorological data in the multidimensional time series dataset is mapped into a deterministic reference power. A dynamic uncertainty quantification mechanism is introduced, and a machine learning model is used to predict the variance of the input meteorological factors. Combined with the instantaneous sensitivity of the physical model chain to each input, the theoretical standard deviation of the physical model is calculated, which is used to quantify the sensitivity of the physical model to the current meteorological conditions. Using the theoretical standard deviation sequence of the above physical model, a physical-guided residual correction dataset is constructed, and the difference between the measured power and the reference power is calculated to obtain the residual sequence. A mixed feature vector and the original NWP meteorological features are constructed. A quantile recurrent neural network is used with the mixed features as input and the residual sequence as the target to learn the residual distribution under different quantiles. By combining the aforementioned physical baseline power with the predicted residual quantiles, the final photovoltaic power probability prediction interval is generated, and the prediction results, guided by physical features and corrected by deep learning, are output.

2. The prediction method according to claim 1, characterized in that, When constructing the physical model chain, calculate the total irradiance of the inclined plane. : ； In the formula, for Total irradiance on the inclined surface of the photovoltaic array at any given time; The component of direct irradiation on the inclined plane; This refers to the irradiance component scattered by the inclined plane. The irradiance component reflected from the sloping ground surface; Calculate the temperature of photovoltaic cells : ； In the formula, The baseline DC power calculated for the physical model; The rated capacity of the photovoltaic array under standard test conditions; The power temperature coefficient; This is the overall system efficiency coefficient, which includes inverter efficiency and line loss.

3. The prediction method according to claim 1, characterized in that, The dynamic uncertainty quantification mechanism, along with the error propagation law, can be broken down into the following three calculation steps: Uncertainty modeling of meteorological factors, utilizing machine learning models for key meteorological factors of NWP, including total irradiance. ,temperature To perform modeling; Using gradient boosting trees (based on weather condition features) Predicting various meteorological input variables conditional variance of : ； In the formula, For the first A meteorological input variable in The variance of the prediction at any given time; For the trained gradient boosting tree model; The input is the state feature vector; the model outputs not only predicted values ​​of meteorological factors. It also outputs the variance of the predicted value. ; The partial derivatives of the physical model's output with respect to each input variable, i.e., the gradient, are calculated using the numerical difference method: ； In the formula, for Power at time t = 1 The physical sensitivity gradient of each variable; It is a tiny numerical perturbation; Total variance of synthetic physics derivation The theoretical standard deviation is obtained by taking the square root: ； In the formula, The theoretical standard deviation derived from the physical model will be passed as a key feature to subsequent steps.

4. The prediction method according to claim 1, characterized in that, The construction of the residual correction dataset specifically involves defining the target residual. : ； In the formula, For true residuals; For actual measured photovoltaic power; Constructing a hybrid feature vector The data is then input into a quantile LSTM network for training, with the objective function being a comprehensive bouncing loss function. : ； In the formula, Set the target quantiles (e.g., 0.025, 0.5, 0.975); The model predicts the first The residual value of the quantile.

5. The prediction method according to claim 1, characterized in that, The formulas for calculating the synthesized physical reference power and the predicted residual quantiles are as follows: ； In the formula, The final confidence level is The power prediction boundary value.

6. The prediction method according to claim 1, characterized in that, Constructing a hybrid feature vector The process is a combination of feature engineering and physical embedding, and involves the following operations in sequence: selection and definition of feature dimensions; concatenation of the above three parts of data along the feature dimensions; and standardization processing.

7. The prediction method according to claim 6, characterized in that, Feature dimension selection and definition: Mixed feature vector At time step It consists of the following three parts of heterogeneous information: Part 1: Deterministic reference power derived from photovoltaic physical model chain as physical reference feature ( ), to provide the main trends and physical lower bounds of power variation; Part Two: Physical confidence characteristics derived from the theoretical standard deviation calculated based on the law of error propagation ( ), to serve as attention weights or gating signals; Part Three: Characteristics of the Primitive Meteorological Environment It provides environmental context information to help neural networks capture nonlinear weather patterns that physical models fail to model.

8. The prediction method according to claim 6, characterized in that, The three data sets mentioned above are concatenated along the feature dimension and then standardized. Assuming time meteorological characteristics Dimensions The construction process is shown in the following formula: ； concatenated feature vectors The total dimension is ; Then process the concatenated vector Each feature dimension in the dataset is Z-score standardized (mean subtracted, standard deviation divided) to make its distribution approximate. Construct a time window sequence This is used as an input sample for the neural network.

9. A hybrid power range prediction system guided by physical uncertainty, characterized in that, include: The data preprocessing module is used to acquire meteorological driving data from numerical weather forecasts and corresponding historical measured data from photovoltaic power plants; perform time zone correction on data from different data sources; convert cumulative quantities such as shortwave radiation into instantaneous quantities and align them with the measured records along the time axis; and remove outlier data. The power calculation module, connected to the data preprocessing module, is used to calculate the solar geometric position, slope irradiance, battery temperature and photoelectric conversion efficiency, and to map the above numerical weather forecast meteorological data into a deterministic reference power. The machine learning module, connected to the power calculation module, calculates the difference between the measured power and the reference power to obtain the residual sequence; constructs a mixed feature vector and the original NWP meteorological features; and uses a quantile recurrent neural network with the mixed features as input and the residual sequence as the target to learn the residual distribution under different quantiles. The output module is connected to the machine learning module. It combines the residual distribution characteristics obtained by machine learning and synthesizes the physical baseline power and the predicted residual quantile to generate the final photovoltaic power probability prediction interval, which is used to output the prediction results guided by physical features and corrected by deep learning. The power supply module is electrically connected to the data preprocessing module, power calculation module, machine learning module, and output module, respectively, and provides power to the data preprocessing module, power calculation module, machine learning module, and output module.

10. A computer-readable storage medium for a hybrid power range prediction system guided by physical uncertainty, characterized in that, The device stores readable computer instructions for performing the steps of the power range prediction method according to claims 1-8.

Images