A photovoltaic power day-ahead prediction method, system, electronic device and storage medium

By constructing a joint historical similar scene retrieval database and a gradient boosting tree model, the problems of data silos and fixed weights in multi-source fusion in traditional photovoltaic forecasting methods are solved, and high-precision photovoltaic power forecasting under complex meteorological conditions is achieved.

CN123073734APending Publication Date: 2026-11-10HANGZHOU JIANJIE WEIJU NEW ENERGY TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611578288.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-13
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Traditional photovoltaic forecasting methods suffer from several drawbacks when dealing with complex weather conditions. These include a lack of quantitative assessment of uncertainties in numerical weather forecasts, a lack of dynamic evolution perception capabilities due to fixed weights in multi-model fusion, and data silos caused by independent modeling at each site. Consequently, power forecasting accuracy is low.

Method used

A joint historical similar scene retrieval database for distributed photovoltaic sites is constructed. By generating a comprehensive divergence index score and a spatial distance index decay function, the contribution weights of cross-site samples are dynamically allocated. Combined with a gradient boosting tree model, fitting residuals and physical restoration are performed to achieve cross-site sharing of meteorological knowledge and adaptive forecasting.

Benefits of technology

It improves the accuracy of photovoltaic power prediction in scenarios of sudden weather changes, solves the technical problems of data silos and fixed weights for multi-source fusion, and realizes adaptive power prediction correction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123073734A_ABST
    Figure CN123073734A_ABST
Patent Text Reader

Abstract

The application relates to a photovoltaic power day-ahead prediction method, system, electronic equipment and computer readable storage medium, a prediction conflict value of two independent day-ahead numerical weather prediction source data in a morning observation window is extracted, positive constraint regression weighted calculation is performed, a comprehensive divergence index score is generated, the total conflict probability and weather confidence of double-source NWP prediction on a whole-day scale are dynamically quantified, the comprehensive divergence index score is combined with historical experience center estimation features extracted by spatial index attenuation retrieval, a pre-trained gradient boosting tree model is input, adaptive tree node path splitting is performed under different meteorological uncertainties, nonlinear clear sky ratio residual errors of a benchmark NWP prediction base are specially used for fitting, technical problems of missing NWP reliability evaluation and fixed multi-source fusion weight in a traditional method are solved, and adaptive correction of photovoltaic day-ahead power prediction accuracy in cloudy and sudden weather change scenes is realized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of power prediction, and in particular to a photovoltaic power day-ahead prediction method, system, electronic device and storage medium. Background Technology

[0002] With the continuous increase in the penetration rate of distributed photovoltaic (PV) power in the power grid, its inherent strong volatility and high randomness pose challenges to the day-ahead dispatching plan and grid absorption of the power system. Traditional day-ahead PV forecasting methods face the following three mutually restrictive technical bottlenecks when dealing with complex weather conditions: First, there is a lack of quantitative assessment of the uncertainty in numerical weather prediction (NWP) forecasts. Existing methods typically use NWP data as indiscriminate feature inputs, failing to establish a mechanism for quantifying weather uncertainty at the model level. During meteorological transition periods such as severe convection, cloudy skies, and the passage of cold fronts, NWP forecast errors exhibit a nonlinear upward trend. If the model cannot dynamically perceive the reliability of the daily NWP forecast, it can lead to nonlinear biases in power predictions under fluctuating weather conditions.

[0003] Secondly, the fixed weights in multi-model fusion lack the ability to perceive dynamic evolution. The industry typically uses fixed weights or static weights updated monthly to fuse multi-source NWP data. This approach severs the physical continuity of forecast errors as they evolve dynamically in time and space. When faced with sudden weather conflicts, it cannot dynamically adjust the trust gating based on the real-time divergence trend of the two-source forecasts, making it difficult for the system to adaptively correct its course.

[0004] Third, independent modeling at each site leads to data silos and a cold start dilemma. Traditional independent modeling paradigms heavily rely on the long-term accumulation of historical samples from a single site. For newly built sites, small-capacity distributed sites, or edge sites with a large amount of missing historical data, directly training high-dimensional nonlinear models can easily lead to overfitting problems. Existing technologies have failed to effectively share common knowledge of the spatiotemporal evolution of cloud conditions among adjacent mature power plants within the region. Summary of the Invention

[0005] This application provides a photovoltaic power day-ahead prediction method, system, electronic device, and computer-readable storage medium to at least address the problem of low accuracy in power prediction results in related technologies.

[0006] In a first aspect, embodiments of this application provide a photovoltaic power day-ahead prediction method, the method comprising: Based on historical measured power generation data and historical multi-source numerical weather forecast data of multiple photovoltaic power stations in the same meteorological region, a joint historical similar scene retrieval database for distributed photovoltaic sites is constructed. Acquire two independent day-ahead numerical weather forecast sources for the forecast date, extract the forecast conflict values ​​within the early morning observation window, perform weighted calculations, and generate a comprehensive divergence index score; Using the forecast meteorological time series characteristics of the predicted day, intra-station and cross-station searches are performed in the joint historical similar scene retrieval database. The contribution weights of cross-station samples are dynamically adjusted through the spatial distance exponential decay function to extract the historical experience center estimation features. The comprehensive divergence index score is combined with the historical experience center estimation features to form an uncertainty-aware feature set. The uncertainty-aware feature set is then input into a pre-trained gradient boosting tree model to fit the residuals and perform physical restoration based on the baseline forecast base, thereby obtaining the day-ahead power prediction curve of the target photovoltaic power plant.

[0007] In some embodiments, a joint historical similarity scenario retrieval library for distributed photovoltaic sites is constructed based on historical measured power generation data and historical multi-source numerical weather forecast data from multiple photovoltaic power stations within the same meteorological region, including: The theoretical clear-sky irradiance of each photovoltaic power station at historical points in time was calculated based on the solar motion trajectory and geometric radiation model. The historical measured power generation data is converted into measured clear sky ratio using the theoretical clear sky irradiance, and the historical multi-source numerical weather forecast data is converted into forecast cloud condition index. The forecast cloud condition index sequence is truncate using a preset sliding window, and the mean, standard deviation and first-order trend slope of the forecast cloud condition index within each sliding window are calculated as the temporal evolution characteristics of the corresponding window. The temporal evolution features extracted within the sliding window are used as retrieval features, and the measured clear sky ratio corresponding to the end time point of the sliding window is used as the effect label. The two are mapped and associated to construct the joint historical similar scene retrieval library.

[0008] In some embodiments, two independent day-ahead numerical weather forecast sources are acquired for the day to be predicted. Forecast conflict values ​​within the early morning observation window are extracted and weighted to generate a comprehensive divergence index score, including: Extract the morning observation window of the morning period of the day to be predicted, and extract the average absolute difference between the first forecast source and the second forecast source in terms of shortwave radiation, total cloud cover and relative humidity in the morning observation window to obtain the forecast conflict component. Using the absolute average error of the actual forecast for the whole day from the first historical forecast source as the target variable, the weight coefficients of each forecast conflict component are learned through positive constraint multiple linear regression. Based on the aforementioned weighting coefficients, each forecast conflict component is weighted and calculated. After normalization and truncation, a comprehensive divergence index score that quantifies the confidence level of the weather throughout the day is output.

[0009] In some embodiments, after generating the comprehensive divergence index score, the method further includes: The slope of divergence evolution is obtained by fitting the comprehensive divergence index score sequence within the morning observation window with a first-order polynomial. The divergence evolution slope and the comprehensive divergence index score are jointly output to the uncertainty perception feature set to characterize the dynamic development trend of weather forecast divergence in the uncertainty perception feature set.

[0010] In some embodiments, dynamically adjusting the contribution weights of cross-site samples using a spatial distance exponential decay function includes: Calculate the Euclidean distance in latitude and longitude between the target photovoltaic power station and the candidate power stations retrieved across stations; Based on the meteorological correlation empirical attenuation coefficient of the target meteorological area and the Euclidean distance of latitude and longitude, the spatial weight of the candidate power station sample is determined by an exponential decay function, and the spatial weight is used as the contribution weight of the cross-station sample. The spatial weight decays exponentially with the increase of the Euclidean distance of latitude and longitude.

[0011] In some embodiments, the path splitting rules of the internal decision tree nodes of the gradient boosting tree model when fitting the residuals include: When the comprehensive divergence index score is lower than the set lower threshold, the decision tree path preferentially splits towards nodes that rely on the features of the baseline numerical weather forecast. When the comprehensive divergence index score is higher than the set upper limit threshold, the decision tree path adaptively splits towards defensive nodes that rely on the estimated features of the historical experience center.

[0012] In some embodiments, inputting the uncertainty-aware feature set into a pre-trained gradient boosting tree model to fit the residuals and perform a physical reconstruction based on the baseline prediction pedestal includes: The gradient boosting tree model is iteratively trained using the difference between the historical measured clear sky ratio and the baseline forecast cloud condition index as the residual learning objective. The clear sky ratio residual prediction value output by the gradient boosting tree model is superimposed on the baseline forecast cloud condition index, and the final clear sky ratio prediction value is obtained after truncation. The day-ahead power prediction curve of the target photovoltaic power station is obtained by multiplying the final clear sky ratio prediction value by the theoretical maximum clear sky power generation of the corresponding power station.

[0013] Secondly, embodiments of this application provide a photovoltaic power day-ahead prediction system, the system comprising: a construction module, a calculation module, an extraction module, and a restoration module, wherein: The construction module is used to build a joint historical similar scene retrieval library for distributed photovoltaic sites based on historical measured power generation data and historical multi-source numerical weather forecast data of multiple photovoltaic power stations in the same meteorological area. The calculation module is used to acquire two independent day-ahead numerical weather forecast source data for the day to be predicted, extract the forecast conflict values ​​within the morning observation window, perform weighted calculations, and generate a comprehensive divergence index score. The extraction module is used to perform intra-station and cross-station searches in the joint historical similar scene retrieval database using the forecast meteorological time series characteristics of the forecast day, and to extract historical experience center estimation features by dynamically adjusting the contribution weight of cross-station samples through the spatial distance exponential decay function. The restoration module is used to combine the comprehensive divergence index score with the historical experience center estimation features into an uncertainty-aware feature set, and input the uncertainty-aware feature set into a pre-trained gradient boosting tree model to fit the residuals and perform physical restoration based on the baseline forecast base to obtain the day-ahead power prediction curve of the target photovoltaic power station.

[0014] Thirdly, embodiments of this application provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described in the first aspect above.

[0015] Fourthly, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect above.

[0016] Compared to related technologies, the photovoltaic power day-ahead prediction method, system, electronic device, and computer-readable storage medium provided in this application primarily address the technical problems of data silos in single-station modeling, lack of reliability assessment in numerical weather prediction, and fixed weights in multi-source fusion. Based on a joint historical similar scenario retrieval database, it achieves cross-station sharing of regional meteorological knowledge and supports cold start for newly established stations; it actively quantifies weather uncertainty by extracting conflict values ​​from dual-source forecasts to generate a comprehensive divergence index score; and it integrates the comprehensive divergence index score with cross-station historical experience features to drive adaptive path splitting of the gradient boosting tree model to specifically fit nonlinear residuals. This solution overcomes the shortcomings of traditional models in their insufficient perception of meteorological uncertainty, achieving adaptive correction of day-ahead power prediction accuracy under sudden weather changes. Attached Figure Description

[0017] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a photovoltaic power day-ahead prediction method according to an embodiment of this application; Figure 2This is a structural block diagram of a photovoltaic power day-ahead prediction system according to an embodiment of this application; Figure 3 This is a schematic diagram of the internal structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the application is described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments provided in this application without inventive effort are within the scope of protection of this application.

[0019] Obviously, the accompanying drawings described below are merely some examples or embodiments of this application. Those skilled in the art can apply this application to other similar scenarios based on these drawings without any inventive effort. Furthermore, it is understood that although the efforts made in this development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this application, any changes to design, manufacturing, or production based on the technical content disclosed in this application are merely conventional technical means and should not be construed as insufficient disclosure of the content of this application.

[0020] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0021] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms “a,” “an,” “an,” “the,” and similar words used in this application do not indicate quantity limitation and may indicate singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units not listed, or may include other steps or units inherent to these processes, methods, products, or devices. The terms “connected,” “linked,” “coupled,” and similar words used in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following objects are in an "or" relationship. The terms "first," "second," and "third" used in this application are merely to distinguish similar objects and do not represent a specific ordering of the objects.

[0022] This embodiment provides a method for predicting photovoltaic power day-ahead. Figure 1 This is a flowchart of a photovoltaic power day-ahead prediction method according to an embodiment of this application, such as... Figure 1 As shown, the process includes the following steps: Step S101: Based on the historical measured power generation data of multiple photovoltaic power stations in the same meteorological area and the historical multi-source numerical weather forecast data, construct a joint historical similar scene retrieval library for distributed photovoltaic sites.

[0023] In this embodiment, in order to break the data silo dilemma of single-station modeling of distributed photovoltaic power plants, an offline cross-station joint historical retrieval database is constructed to aggregate the historical data of various heterogeneous photovoltaic power plants in the same meteorological area in a matrix manner.

[0024] First, standardized alignment and access of multi-source heterogeneous spatiotemporal data is performed. Historical 15-minute granular measured power sequences, rated installed capacity, and latitude and longitude coordinates of N photovoltaic power plants in the target area are collected. Simultaneously, dual-source NWP grid data from the first historical forecast source (such as the ICON model) and the second forecast source (such as the GFS model) are accessed. The two forecast grids are aligned and interpolated in the spatiotemporal dimension through timestamp matching to an equal-interval time series consistent with the measured sequences.

[0025] In real-world scenarios, due to the varying installed capacity and geographical locations of different photovoltaic power stations, direct comparison of absolute power values ​​is impossible. This embodiment employs physical-astronomical base decoupling and clear-sky ratio spatial mapping. Specifically, it calls astrophysical computing components to calculate the theoretical clear-sky irradiance, solar zenith angle, and azimuth angle of each photovoltaic power station at historical points in time, representing pure astrophysical characteristics. Only power generation periods with theoretical clear-sky irradiance exceeding a preset effective threshold (e.g., 20 W / m²) are retained. Historical measured power generation data is transformed into dimensionless measured clear-sky ratios, while historical multi-source numerical weather forecast data is converted into forecast cloud condition indices.

[0026] The specific mapping formula is as follows: p_theo = capacity × (ghi_cs / G_stc) kc_real = clip(power / (p_theo + ε),0, 1.3) cloud_index_icon = clip(SW_radiation_icon / ghi_cs, 0, 1.3) In the formula, p_theo represents the theoretical maximum power generation under clear skies, capacity represents the rated installed capacity of the power station, ghi_cs represents the theoretical clear skies irradiance calculated by the geometric radiation model, G_stc represents the baseline irradiance constant under standard test conditions, ε represents the regularization minimum constant to prevent the denominator from being zero, the upper limit of the clip function is 1.3 to be compatible with short-term inverter overload and cloud edge enhancement effects, kc_real is the measured clear skies ratio, and cloud_index_icon is the forecast cloud condition index of the baseline first forecast source.

[0027] Furthermore, after obtaining the aforementioned dimensionless features, the forecast cloud condition index sequences in the historical training set of all power stations are extracted in time series using a preset sliding window (e.g., a length of 12, corresponding to 3 hours). The mean, standard deviation, and first-order trend slope of the sequences within each sliding window are calculated to construct a three-dimensional time series evolution feature vector as the retrieval feature. The measured clear sky ratio corresponding to the end of the sliding window is used as the effect label. Based on the hash serialization mapping association of power station identifiers, a historical similar scene retrieval library supporting joint retrieval of multiple stations is constructed.

[0028] Step S102: Obtain two independent day-ahead numerical weather forecast source data for the day to be predicted, extract the forecast conflict values ​​within the early morning observation window, perform weighted calculations, and generate a comprehensive divergence index score.

[0029] In this embodiment, two independent NWP grid data for the next day are automatically retrieved at a fixed pre-day scheduling time, and a dual-source NWP divergence index (DI) data-driven engine is established to proactively quantify the weather uncertainty of the next day.

[0030] Specifically, the observation window for the early morning period following the predicted power generation time is selected (corresponding to the period from 06:00 to 08:00). It should be noted that this period coincides with the sunrise initiation phase of a thermodynamic abrupt change in the atmospheric boundary layer, characterized by a low solar altitude angle and a high atmospheric optical mass number. This amplifies minor meteorological disturbances such as localized radiation fog, morning cloud fragmentation, and dew formation due to supersaturation. Due to differences in boundary layer parameterization schemes among forecast models, significant phase differentiation will occur in the forecast flow within this window. The divergence of this initial phase can serve as a forward-looking assessment indicator of the difficulty of nonlinear forecasting of all-day weather systems.

[0031] Furthermore, during this early morning observation window, the average absolute difference between the first and second forecast sources on core meteorological elements is extracted as the forecast conflict component, specifically including the absolute difference in surface shortwave radiation forecast (di_rad), the absolute difference in total cloud cover forecast (di_cloud), and the absolute difference in near-surface relative humidity forecast (di_hum).

[0032] To assign reasonable physical weights to each conflicting forecast component, the absolute average error of the historical first forecast source for the entire day is used as the objective variable. Weight coefficients are automatically learned offline using positive-constrained multiple linear regression. The objective function is defined as minimizing the sum of squared residuals (RSS) of the entire historical sample. RSS = ∑(nwp_err[i] - (w_rad × di_rad[i]+ w_cloud × di_cloud[i] +w_hum × di_hum[i]))² Simultaneously, positive constraints are applied during the solution process: w_rad ≥ 0, w_cloud ≥ 0, w_hum ≥ 0. Here, nwp_err[i] represents the absolute average error of the actual forecast from the first forecast source on the i-th historical day. The weighted coefficients w_rad, w_cloud, and w_hum, output through this convex optimization iteration, can automatically characterize the local meteorological characteristics of the target area, avoiding subjective human bias.

[0033] Based on the aforementioned weighting coefficients, each forecast conflict component is weighted and calculated. After normalization and two-end truncation, a comprehensive divergence index score di_score, quantifying the confidence level of the weather throughout the day, is output: di_score = clip((w_rad × di_rad + w_cloud × di_cloud + w_hum × di_hum) / S_scale, 0, 1) In the formula, S_scale is the normalization scale constant, which is extracted and calibrated from the preset quantile (such as the 90th percentile) of the divergence of the full sample distribution of historical extreme weather, and is used to map the unbounded absolute value of divergence to the standard probability confidence interval.

[0034] Furthermore, this embodiment also obtains the divergence evolution slope di_slope by fitting the time series sequence of the comprehensive divergence index score within the morning observation window using a first-order polynomial. A slope greater than zero indicates a continuous widening of the divergence between the two-source forecasts, suggesting a sharp increase in weather uncertainty; a slope less than zero indicates a narrowing of the divergence. This slope feature, along with the comprehensive divergence index score, is output to the subsequent feature set.

[0035] Step S103: Using the forecast meteorological time series characteristics of the predicted day, perform intra-station and cross-station searches in the joint historical similar scene retrieval database, and dynamically adjust the contribution weight of cross-station samples through the spatial distance exponential decay function to extract the historical experience center estimation features.

[0036] In this embodiment, cross-site time-series evolution experience retrieval is performed concurrently during the day-ahead online forecast phase. Sliding window features of the next-day forecast cloud condition index are extracted from the forecast site and used as a query vector. Simultaneously, in a joint historical similarity scenario retrieval database, the most similar historical time periods are searched in the historical database of this site and the historical databases of candidate adjacent power stations within the region.

[0037] To effectively integrate search results from different locations, a spatial distance exponential decay function is introduced to dynamically calculate the contribution weight of cross-site samples, as shown in the following formula: dist = √[(lat_A - lat_B)² + (lon_A - lon_B)²] w_space = exp(-dist / λ) In the formula, dist represents the Euclidean distance in latitude and longitude between the target photovoltaic power station and the candidate power stations in the cross-site retrieval; λ represents the empirical attenuation coefficient of meteorological correlation of the target meteorological region, which is calibrated by the spatial correlation of regional meteorological evolution. This function makes the spatial weight of the candidate power station samples decrease exponentially with the increase of the Euclidean distance in latitude and longitude, ensuring the geophysical rationality of meteorological feature fusion.

[0038] Finally, after performing matrix-level weighted fusion on the retrieved set based on the calculated spatial weights, the mean of the clear sky ratio of historical similar scenes (anen_mean) and the corresponding uncertainty measure (anen_std) are extracted and used as the historical experience center estimation features.

[0039] Understandably, this cross-site retrieval mechanism also supports cold start scenarios. When historical data for a newly established target site is scarce, it can supplement the retrieval by using the historical database of mature sites in the climate-related region, thereby directly obtaining effective historical prior knowledge and enabling the newly established site to output prediction results with usable accuracy in the early stages.

[0040] Step S104: Combine the comprehensive divergence index score with the historical experience center estimation features into an uncertainty-aware feature set. Input the uncertainty-aware feature set into the pre-trained gradient boosting tree model to fit the residuals and perform physical restoration based on the baseline forecast base to obtain the day-ahead power prediction curve of the target photovoltaic power station.

[0041] In this embodiment, a residual learning architecture anchored to a physical base is employed. Instead of directly fitting the absolute power value using a machine learning model, it specifically predicts and hedges the nonlinear bias of NWP in the clear sky ratio dimension. The residual learning objective is defined as the difference between the historically measured clear sky ratio and the baseline forecast cloud condition index: kc_residual = kc_real - cloud_index_icon During the offline training phase of the model, the simplified and refined physical features (theoretical clear-sky irradiance, solar position features, and temporal cyclic encoding), the features output by the DI engine (di_score, di_slope), and the historical experience center estimation features (anen_mean, anen_std) output by cross-site retrieval are assembled into a global feature matrix. This simplified anti-dilution combination strategy aims to maintain the gradient splitting focus of the tree model in a low-dimensional space. Subsequently, iterative training and early stopping calibration of the gradient boosting tree (such as LightGBM) model are performed using the minimum mean squared error (MSE) as the loss function.

[0042] During online application, the online forecast feature matrix of the corresponding forecast date is input into the pre-trained gradient boosting tree model. When fitting the residuals, the comprehensive divergence index score di_score is used as a high-dimensional decision feature to implicitly perform confidence-gated routing and path splitting: when the comprehensive divergence index score is lower than a set lower threshold (e.g., 0.20), it indicates that the current weather system is stable and the NWP forecast confidence is high, and the decision tree path preferentially splits towards physical fine-tuning points that rely on the baseline numerical weather forecast feature (cloud condition index); when the comprehensive divergence index score is higher than a set upper threshold (e.g., 0.50), it indicates that there is a risk of drastic weather change, and the decision tree path automatically and adaptively flows to defensive nodes that heavily rely on the historical experience center estimation feature (anen_mean) for splitting and combination; in the transition zone between the two, the tree model automatically interpolates the routing based on the feature distribution.

[0043] After the model outputs the clear sky ratio residual prediction value delta_kc_pred for each time point (e.g., 96 15-minute nodes) the next day, it is superimposed on the baseline forecast cloud condition index through the physical constraint inversion formula. After truncation, the final clear sky ratio prediction value kc_pred_final is obtained, and finally multiplied by the theoretical maximum clear sky power generation of the corresponding power station to restore the power prediction curve: kc_pred_final = clip(cloud_index_icon + delta_kc_pred, 0, 1.3) P_pred =kc_pred_final × p_theo.

[0044] Finally, after obtaining the day-ahead power prediction curve P_pred for the target photovoltaic power plant, the data is formatted and reported according to the standard specifications of the power dispatch automation system. Through the above decoupled design driven by both physical and data aspects, the method maintains the smoothness of the prediction trajectory and achieves adaptive error suppression when dealing with various strong disturbance weather scenarios.

[0045] Through steps S101 to S104 above, the method provided in this application mainly addresses the technical problems of data silos in single-station modeling, lack of reliability assessment for numerical weather prediction, and fixed weights in multi-source fusion. This method, based on a joint historical similarity scenario retrieval database, breaks down data silos, enabling cross-station sharing of regional meteorological knowledge and supporting cold start for new stations. It actively quantifies weather forecast confidence by extracting conflict values ​​from dual-source forecasts to generate a comprehensive divergence index score. By fusing the comprehensive divergence index score with cross-station historical experience features, it drives the gradient boosting tree model to perform adaptive path splitting to specifically fit nonlinear residuals. This approach compensates for the shortcomings of traditional models in perceiving meteorological uncertainties, achieving adaptive correction of day-ahead power prediction accuracy under sudden weather changes.

[0046] On the other hand, this embodiment also provides a photovoltaic power day-ahead prediction system. Figure 2 This is a structural block diagram of a photovoltaic power day-ahead prediction system according to an embodiment of this application, as shown below. Figure 2 As shown, the system includes: a construction module 20, a calculation module 21, an extraction module 22, and a restoration module 23, wherein: The construction module 20 is used to build a joint historical similar scene retrieval library for distributed photovoltaic sites based on the historical measured power generation data and historical multi-source numerical weather forecast data of multiple photovoltaic power stations in the same meteorological area. The calculation module 21 is used to acquire two independent day-ahead numerical weather forecast source data for the day to be predicted, extract the forecast conflict values ​​within the morning observation window, perform weighted calculations, and generate a comprehensive divergence index score. The extraction module 22 is used to perform intra-station and cross-station searches in the joint historical similar scene retrieval database using the forecast meteorological time series characteristics of the forecast day, and to extract historical experience center estimation features by dynamically adjusting the contribution weight of cross-station samples through the spatial distance exponential decay function. The restoration module 23 is used to combine the comprehensive divergence index score with the historical experience center estimation features into an uncertainty-aware feature set, input the uncertainty-aware feature set into a pre-trained gradient boosting tree model for fitting residuals and physical restoration based on the baseline forecast base, and obtain the day-ahead power prediction curve of the target photovoltaic power station.

[0047] The system provided in this application primarily addresses the technical problems of data silos in single-station modeling, lack of reliability assessment in numerical weather prediction, and fixed weights in multi-source fusion. It breaks down single-station limitations and shares regional knowledge through a construction module; it extracts conflict values ​​from dual-source forecasts using a calculation module to proactively quantify weather uncertainty; it introduces a spatial distance exponential decay function through an extraction module to dynamically allocate cross-station sample weights to extract empirical features; and finally, a restoration module drives the gradient boosting tree model to adaptively split the path, accurately fit the residuals, and invert the restored power curve. This solution effectively improves the accuracy of day-ahead photovoltaic power prediction under strongly disturbed meteorological scenarios.

[0048] In one embodiment, Figure 3 This is a schematic diagram of the internal structure of an electronic device according to an embodiment of this application. For example... Figure 3 As shown, an electronic device is provided, which can be an in-vehicle computing unit or a cloud server. The electronic device includes a processor, a network interface, internal memory, and non-volatile memory connected via an internal bus. The non-volatile memory stores parameters related to the operating system, computer programs, and a large language model. The processor provides computing power to execute deep neural network inference and control algorithm calculations. When the processor executes the stored computer program, it implements a photovoltaic power day-ahead prediction method as described in any of the above embodiments.

[0049] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the electronic device to which the present application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0050] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by hardware related to computer program instructions. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0051] The above embodiments merely illustrate several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for predicting photovoltaic power day-ahead, characterized in that, The method includes: Based on historical measured power generation data and historical multi-source numerical weather forecast data of multiple photovoltaic power stations in the same meteorological region, a joint historical similar scene retrieval database for distributed photovoltaic sites is constructed. Acquire two independent day-ahead numerical weather forecast sources for the forecast date, extract the forecast conflict values ​​within the early morning observation window, perform weighted calculations, and generate a comprehensive divergence index score; Using the forecast meteorological time series characteristics of the predicted day, intra-station and cross-station searches are performed in the joint historical similar scene retrieval database. The contribution weights of cross-station samples are dynamically adjusted through the spatial distance exponential decay function to extract the historical experience center estimation features. The comprehensive divergence index score is combined with the historical experience center estimation features to form an uncertainty-aware feature set. The uncertainty-aware feature set is then input into a pre-trained gradient boosting tree model to fit the residuals and perform physical restoration based on the baseline forecast base, thereby obtaining the day-ahead power prediction curve of the target photovoltaic power plant.

2. The method according to claim 1, characterized in that, Based on historical measured power generation data and historical multi-source numerical weather prediction data from multiple photovoltaic power stations within the same meteorological region, a joint historical similarity scenario retrieval database for distributed photovoltaic sites is constructed, including: The theoretical clear-sky irradiance of each photovoltaic power station at historical points in time was calculated based on the solar motion trajectory and geometric radiation model. The historical measured power generation data is converted into measured clear sky ratio using the theoretical clear sky irradiance, and the historical multi-source numerical weather forecast data is converted into forecast cloud condition index. The forecast cloud condition index sequence is truncate using a preset sliding window, and the mean, standard deviation and first-order trend slope of the forecast cloud condition index within each sliding window are calculated as the temporal evolution characteristics of the corresponding window. The temporal evolution features extracted within the sliding window are used as retrieval features, and the measured clear sky ratio corresponding to the end time point of the sliding window is used as the effect label. The two are mapped and associated to construct the joint historical similar scene retrieval library.

3. The method according to claim 1, characterized in that, Obtain data from two independent day-ahead numerical weather prediction sources for the forecast date, extract conflicting forecast values ​​within the early morning observation window, perform weighted calculations, and generate a comprehensive divergence index score, including: Extract the morning observation window of the morning period of the day to be predicted, and extract the average absolute difference between the first forecast source and the second forecast source in terms of shortwave radiation, total cloud cover and relative humidity in the morning observation window to obtain the forecast conflict component. Using the absolute average error of the actual forecast for the whole day from the first historical forecast source as the target variable, the weight coefficients of each forecast conflict component are learned through positive constraint multiple linear regression. Based on the aforementioned weighting coefficients, each forecast conflict component is weighted and calculated. After normalization and truncation, a comprehensive divergence index score that quantifies the confidence level of the weather throughout the day is output.

4. The method according to claim 3, characterized in that, After generating the comprehensive divergence index score, the method further includes: The slope of divergence evolution is obtained by fitting the comprehensive divergence index score sequence within the morning observation window with a first-order polynomial. The divergence evolution slope and the comprehensive divergence index score are jointly output to the uncertainty perception feature set to characterize the dynamic development trend of weather forecast divergence in the uncertainty perception feature set.

5. The method according to claim 1, characterized in that, The contribution weights of cross-site samples are dynamically adjusted using a spatial distance exponential decay function, including: Calculate the Euclidean distance in latitude and longitude between the target photovoltaic power station and the candidate power stations retrieved across stations; Based on the meteorological correlation empirical attenuation coefficient of the target meteorological area and the Euclidean distance of latitude and longitude, the spatial weight of the candidate power station sample is determined by an exponential decay function, and the spatial weight is used as the contribution weight of the cross-station sample. The spatial weight decays exponentially with the increase of the Euclidean distance of latitude and longitude.

6. The method according to claim 1, characterized in that, When fitting the residuals, the path splitting rules of the decision tree nodes in the gradient boosting tree model include: When the comprehensive divergence index score is lower than the set lower threshold, the decision tree path preferentially splits towards nodes that rely on the features of the baseline numerical weather forecast. When the comprehensive divergence index score is higher than the set upper limit threshold, the decision tree path adaptively splits towards defensive nodes that rely on the estimated features of the historical experience center.

7. The method according to claim 2, characterized in that, The process of inputting the uncertainty-aware feature set into a pre-trained gradient boosting tree model for fitting residuals and physical reconstruction based on the baseline prediction base includes: The gradient boosting tree model is iteratively trained using the difference between the historical measured clear sky ratio and the baseline forecast cloud condition index as the residual learning objective. The clear sky ratio residual prediction value output by the gradient boosting tree model is superimposed on the baseline forecast cloud condition index, and the final clear sky ratio prediction value is obtained after truncation. The day-ahead power prediction curve of the target photovoltaic power station is obtained by multiplying the final clear sky ratio prediction value by the theoretical maximum clear sky power generation of the corresponding power station.

8. A photovoltaic power day-ahead prediction system, characterized in that, The system includes: a construction module, a calculation module, an extraction module, and a restoration module, wherein: The construction module is used to build a joint historical similar scene retrieval library for distributed photovoltaic sites based on historical measured power generation data and historical multi-source numerical weather forecast data of multiple photovoltaic power stations in the same meteorological area. The calculation module is used to acquire two independent day-ahead numerical weather forecast source data for the day to be predicted, extract the forecast conflict values ​​within the morning observation window, perform weighted calculations, and generate a comprehensive divergence index score. The extraction module is used to perform intra-station and cross-station searches in the joint historical similar scene retrieval database using the forecast meteorological time series characteristics of the forecast day, and to extract historical experience center estimation features by dynamically adjusting the contribution weight of cross-station samples through the spatial distance exponential decay function. The restoration module is used to combine the comprehensive divergence index score with the historical experience center estimation features into an uncertainty-aware feature set, and input the uncertainty-aware feature set into a pre-trained gradient boosting tree model to fit the residuals and perform physical restoration based on the baseline forecast base to obtain the day-ahead power prediction curve of the target photovoltaic power station.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.