An ultra-short-term solar irradiance rolling prediction method and system

Abstract

This invention discloses a method and system for ultra-short-term rolling solar irradiance forecasts. The method first acquires historical measured sequences of solar irradiance and meteorological observation elements, and then obtains model forecast sequences of solar irradiance and meteorological elements based on the Huafeng numerical model. Next, for each forecast period, an independent LightGBM regression model is established to predict the recursive forecast value of solar irradiance for that forecast period. The recursive forecast values ​​are then integrated in chronological order to obtain the solar irradiance forecast sequence. The previously optimized recursive forecast values ​​replace the original solar irradiance model forecast values, driving the LightGBM model for the next forecast period. This invention efficiently fuses low-frequency model fields with high-frequency observations, constructing a temporal dependency relationship between adjacent forecast periods, significantly improving the accuracy of multi-step forecasts.

Description

Technical Field

[0001] This invention relates to the field of meteorological forecasting technology, and in particular to a method and system for ultra-short-term rolling solar irradiance forecasting. Background Technology

[0002] As a core component of new energy power systems, solar photovoltaic (PV) power generation presents significant challenges to the safe and stable operation of power systems and the efficient absorption of new energy due to its inherent intermittency and volatility. Accurate 0-4 hour ultra-short-term solar irradiance forecasts are the core foundation for PV power prediction, directly determining the grid's capacity to absorb PV new energy and the safety of its dispatch operations. They are also a key technical support for optimizing grid dispatch plans and ensuring the smooth operation of the electricity spot market.

[0003] Currently, ultra-short-term solar irradiance forecasting technologies mainly fall into three major categories, and each of these technologies has significant technical shortcomings and application bottlenecks in practical operational applications:

[0004] The first category is cloud trajectory extrapolation methods based on ground-based all-sky imagers or geostationary satellite cloud images. These methods can capture short-term changes in irradiance by extrapolating the movement trajectory of cloud clusters, but their effective forecast lead time is usually limited to 0-3 hours, and they are highly dependent on expensive ground-based observation hardware or high-frequency satellite remote sensing data. The operational deployment and maintenance costs are high, and it is difficult to promote and apply them on a large scale.

[0005] The second category is the direct forecasting method of operational numerical weather prediction (NWP) models. Domestic operational regional high-resolution models, represented by the Huafeng numerical model, use a twice-daily reporting frequency (08:00 and 20:00) to output irradiance forecasts every 15 minutes for the next 10 days. However, this method has two major drawbacks: First, limited by computational resources and data assimilation cycles, the model suffers from a technical defect of "low-frequency reporting and high-frequency output," resulting in a significant lag in background field information updates. It cannot promptly reflect the rapid evolution of local weather systems and is slow to respond to sudden changes in irradiance caused by the formation, dissipation, and movement of local cloud clusters. Forecast accuracy decreases significantly with increasing forecast lead time. Second, the model exhibits significant systematic forecast bias. Validation results from multiple typical stations show that the model's summer irradiance forecast values ​​are significantly lower than expected, failing to meet the high-precision operational requirements of 0-4 hour lead time for ultra-short-term photovoltaic forecasts.

[0006] The third category is NWP bias correction forecasting methods based on statistical models or machine learning. These methods establish a mapping relationship between NWP output and measured ground irradiance using historical data to correct biases in model forecasts. However, relying solely on historical observation sequences for pure time-series extrapolation (such as ARIMA and LSTM models), while capturing the short-term persistence characteristics of irradiance, loses trend information on key physical elements in NWP such as future cloud cover, precipitation, and temperature and humidity profiles. This results in low forecast accuracy, with a sharp decline in forecast capability during weather transitions. Furthermore, existing schemes employ a "one-input, multi-output" strategy when performing multi-step forecasts, leading to a significant decrease in forecast accuracy after the second time step, making it difficult to meet the high-precision operational requirements for full-time forecasts from 0 to 4 hours. Summary of the Invention

[0007] Purpose of the invention: The purpose of this invention is to provide a method and system for ultra-short-term solar irradiance rolling forecast, which effectively integrates low-frequency NWP model field trend information with real-time evolution information from high-frequency ground observations, maintains temporal consistency between multi-step forecasts, and improves forecast accuracy.

[0008] Technical solution: The ultra-short-term solar irradiance rolling forecast method of the present invention includes the following steps:

[0009] Obtain historical measured sequences of solar irradiance and historical measured sequences of meteorological observation elements;

[0010] The solar irradiance model forecast sequence and the meteorological forecast element model forecast sequence were obtained based on the Huafeng numerical model.

[0011] For each forecast time, an independent LightGBM regression model is established to predict the recursive forecast value of solar irradiance for that forecast time. The recursive forecast values ​​are then integrated in chronological order to obtain the solar irradiance forecast sequence.

[0012] For the k-th forecast time, the output of the corresponding LightGBM regression model is the solar irradiance model forecast value SR_pred(T+k) for the k-th forecast time, where T is the current time. The input features of the LightGBM regression model include the historical measured sequence of solar irradiance, the historical measured sequence of meteorological observation elements, the model forecast value of meteorological forecast elements for the k-th forecast time, and SR_pred(T+k-1). Among them, the model forecast value of meteorological forecast elements is taken from the model forecast sequence of meteorological forecast elements. When k=0, SR_pred(T+k-1) is replaced by the total radiative radiation SR(T) of the Chinese wind at time T in the solar irradiance model forecast sequence.

[0013] Furthermore, each 15-minute interval within the next 0 to 4 hours is taken as a forecast interval, k=0, 1, 2,..., 16.

[0014] Furthermore, meteorological observation elements include at least one of instantaneous wind speed, instantaneous wind direction, hourly cumulative precipitation, and horizontal visibility.

[0015] Furthermore, meteorological forecast elements include at least one of the following: total cloud cover, low cloud cover, middle cloud cover, high cloud cover, cumulative precipitation, temperature, humidity, wind speed, wind direction, gusts, and air pressure.

[0016] Furthermore, the method also includes aligning the acquired historical measured sequences of solar irradiance, historical measured sequences of meteorological observation elements, solar irradiance model forecast sequences, and meteorological forecast element model forecast sequences according to the time dimension, and cleaning the data based on physical rationality.

[0017] Furthermore, the loss function of the LightGBM regression model is the mean square error between the recursive predicted value and the measured irradiance.

[0018] Furthermore, when the forecast time falls at night, the solar irradiance model forecast value for that forecast time is forced to be 0.

[0019] The present invention provides an ultra-short-term solar irradiance rolling forecast system, comprising:

[0020] The measured data acquisition unit is used to acquire historical measured sequences of solar irradiance and historical measured sequences of meteorological observation elements.

[0021] The numerical model forecast data acquisition unit is used to acquire solar irradiance model forecast sequences and meteorological forecast element model forecast sequences based on the Huafeng numerical model.

[0022] The recursive forecast unit is used to independently build a LightGBM regression model for each forecast time, forecast the recursive forecast value of solar irradiance for that forecast time, and integrate the recursive forecast values ​​in chronological order to obtain a solar irradiance forecast sequence.

[0023] For the k-th forecast time, the output of the corresponding LightGBM regression model is the solar irradiance model forecast value SR_pred(T+k) for the k-th forecast time. The input features of the LightGBM regression model include the historical measured sequence of solar irradiance, the historical measured sequence of meteorological observation elements, the model forecast value of meteorological forecast elements for the k-th forecast time, and SR_pred(T+k-1). Among them, the model forecast value of meteorological forecast elements is taken from the model forecast sequence of meteorological forecast elements. When k=0, SR_pred(T+k-1) is replaced by the total radiative radiation SR(T) of the Chinese wind at time T in the model forecast sequence of solar irradiance.

[0024] The electronic device of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded onto the processor, it implements the ultra-short-term solar irradiance rolling forecast method.

[0025] The computer program product of the present invention includes a computer program that, when executed by a processor, implements the ultra-short-term solar irradiance rolling forecast method.

[0026] Beneficial Effects: This invention introduces a recursive forecasting strategy to establish temporal dependencies between adjacent forecast times. It replaces the original model's total irradiance forecast with the optimized recursive forecast value from the previous time period, driving the LightGBM model to make predictions for the next time period, forming a recursive forecast chain. Simultaneously, it integrates finely selected numerical model forecast features with high-frequency ground observation features, independently training a dedicated model for each forecast time period. This allows for the independent generation of high-precision solar irradiance forecasts for the next 0-4 hours in 15-minute increments without requiring high-frequency updates to the NWP (Natural Power Surface). Compared with existing technologies, the advantages of this invention are:

[0027] (1) This invention proposes an ultra-short-term irradiance rolling forecast method that does not rely on high-frequency NWP updates. This method uses low-frequency NWP as a "trend-leading field" and high-frequency observation autoregressive term as a "real-time correction field". The two are organically integrated through a recursive forecast chain to independently produce forecasts, thus solving the technical problem that low-frequency NWP data cannot support the high-frequency operational forecast requirements.

[0028] (2) This invention proposes a recursive forecasting strategy, which constructs a temporal dependency between adjacent forecast times by replacing the original total radiation forecast value of the Chinese wind with the preceding recursive forecast value. This strategy keeps the RMSE and MAE indices of each forecast time stable, and fundamentally solves the core problem of the significant decay of forecast accuracy after the second time, which is common in traditional one-input, multi-step output forecasting strategies.

[0029] (3) This invention systematically screens numerical model forecast variables and other observation elements. Through physical correlation analysis and multiple combination experiments, it retains a subset of variables that have clear physical indication significance for radiation forecasting. This ensures the model's ability to capture key physical processes while avoiding noise introduced by redundant variables, thus balancing accuracy and computational efficiency.

[0030] (4) This invention uses day and night marking to force the night forecast results to zero, thus avoiding non-physical forecast results with non-zero nighttime irradiance; by performing physical processing synchronously in the recursive chain, it ensures that the recursive error is not amplified by non-physical values. The forecast results can be directly connected to business systems such as photovoltaic power prediction and power grid dispatch, and have strong engineering practicality.

[0031] (5) This invention only requires conventional meteorological observation data and existing twice-daily NWP data to operate, without the need for additional hardware investment or model upgrades. It has low operational deployment costs and strong scalability. The final selected subset of model variables consists of about 5 core variables, which makes the feature engineering complexity controllable and balances forecast accuracy and computational resource requirements. Attached Figure Description

[0032] Figure 1 This is a flowchart of the forecasting method according to an embodiment of the present invention.

[0033] Figure 2 This is a schematic diagram illustrating the temporal correspondence between feature inputs and prediction labels in the recursive prediction architecture of this invention.

[0034] Figure 3 The following is a combined analysis of the following: (a) scatter plot comparison of the recursive forecast values ​​for the next 15 minutes (Lead=1) with the observed winds and winds, (b) error distribution, (c) complete test set time series, and (d) combined time series analysis of the last two days of the test set, according to an embodiment of the present invention.

[0035] Figure 4 The following is a combination analysis of the recursive forecast values ​​for the next 4 hours (Lead=16) of this invention with the observations of the Huafeng (a) scatter plot, (b) error distribution, (c) complete test set time series, and (d) combined time series of the last two days of the test set.

[0036] Figure 5 This is a combined analysis diagram of the recursive forecast values ​​for the next 15 minutes (Lead=1) of this invention with the (a) time series, (b) 15-minute MAPE comparison, and (c) daily average MAPE comparison of the observed winds and winds.

[0037] Figure 6 This is a combined analysis chart of the recursive forecast values ​​for the next 4 hours (Lead=16) and the observed (a) time series, (b) 15-minute MAPE comparison, and (c) daily average MAPE comparison of the embodiment of the present invention. Detailed Implementation

[0038] In this invention, solar irradiance and total radiation both refer to shortwave radiation received by a horizontal surface (unit: W / m²). They are numerically equivalent, differing only in the naming conventions of the data sources. The numerical model (Huafeng) uses the meteorological standard variable name "total radiation," while ground observations in photovoltaic operations often refer to it as "irradiance."

[0039] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0040] like Figure 1 As shown, the ultra-short-term solar irradiance rolling forecast method includes the following steps.

[0041] Step 1: Multi-source data acquisition and preprocessing

[0042] (1) Collect measured irradiance data from ground stations. In this embodiment, the historical measured sequence of surface solar irradiance of the target station is obtained every 15 minutes. Quality control is performed and outliers are removed.

[0043] (2) Collect other meteorological observation data from ground stations. In this embodiment, the historical measured sequence of other meteorological observation elements of the target station is obtained every 15 minutes, including but not limited to air pressure (PRS), sea level pressure (PRS_Sea), air temperature (TEM), dew point temperature (DPT), relative humidity (RHU), water vapor pressure (VAP), instantaneous wind speed (WIN_S_INST), instantaneous wind direction (WIN_D_INST), hourly cumulative precipitation (PRE), horizontal visibility (VIS_HOR_1MI, VIS_HOR_10MI), etc., and quality control is performed to remove outliers.

[0044] (3) Data from the Huafeng numerical model were collected. In this embodiment, the latest forecast data from the model at 08:00 and 20:00 daily were selected, and the element fields with a forecast frequency of 15 minutes and a forecast lead time covering the next 10 days were extracted. Core forecast variables were extracted for the target site, including total radiation (SR), total cloud cover (TCC), low cloud cover (LCC), medium cloud cover (MCC), high cloud cover (HCC), cumulative precipitation in the previous 15 minutes (PRE_15m), multi-layer temperature (TEM), multi-layer humidity (RHU), multi-layer wind speed (WS), multi-layer wind direction (WD), gust (GUST), and multi-layer pressure (PRS). Among them, the multi-layer elements include 8 height layers (10m, 30m, 50m, 70m, 100m, 120m, 150m, 200m). The grid averaging method is used to extract the time series of meteorological elements at station locations, enabling spatial matching between the model field and the stations.

[0045] Step 2: Data Cleaning and Alignment

[0046] (1) Time dimension alignment: The numerical model forecast data, the station measured irradiance data, and other observation element data are aligned according to the time dimension of 15 minutes. Abnormal intervals in the numerical model forecast data are detected and marked; abnormal intervals detected in the observation data (such as those exceeding the normal fluctuation range or having long continuous missing periods) are marked.

[0047] (2) Data quality control: Physical rationality checks are performed on various elements, including: replacing wind direction-related variables with missing values ​​if they exceed the range of [0, 360]; and handling extremely outliers (such as those greater than 360). The missing values ​​are replaced with the filled values. Missing values ​​are filled using linear interpolation; data gaps exceeding 12 hours are removed. The final result is a clean data table on a unified timeline.

[0048] Step 3: Construction of Training and Test Sets

[0049] (1) Sample structure construction: The merged clean data table is processed using a sliding window. The input features of a single training sample include: the total radiation forecast (SR) value of the numerical model, a subset of other meteorological forecast elements of the numerical model, the measured irradiance values ​​of the past n time periods, and the values ​​of other meteorological observation elements of the past i time periods. The training label of a single sample is the measured irradiance value of the station at that time period.

[0050] (2) Dataset partitioning: The sample set is partitioned using a time-ordered partitioning method, without using a random shuffling strategy. This partitioning method ensures that both the training set and the test set cover similar seasonal characteristics and meteorological change patterns, while strictly guaranteeing the continuity of the irradiance time-series data, thus fundamentally avoiding the future data leakage problems that may be caused by random partitioning.

[0051] Step 4: Time-series independent modeling and hyperparameter optimization based on recursive forecast architecture

[0052] (1) Recursive forecast architecture design: In this embodiment, for each 15-minute forecast time (denoted as the k-th time, k=0, 1, 2, ..., 16; where k=0 is the current time, and k=1 to 16 are the next 15 minutes to 240 minutes) within the next 0-4 hours, a LightGBM regression model M is trained independently. k Model M k The input value for the total radiation SR of the Chinese wind is determined using a recursive strategy: when k=0, the original predicted total radiation value SR(T) of the Chinese wind is used as the input; when k≥1, the preceding model M is used. k-1 The recursive forecast value SR_pred(T+k-1) for time t+k-1 replaces the original total radiation forecast value SR(T+k) for the Chinese winds as the SR input to the model. For other numerical models, the original forecast values ​​for the corresponding time T+k are always used for other forecast variables. Historical measured irradiance and other observational features maintain a fixed historical observation window, always selecting the measured values ​​from the few times closest to the forecast start time. This forms the following sequence: "Original Chinese Winds SR — M0 — Predicted SR0 — M1 — Predicted SR1 — M2 — Predicted SR2 — ... — M 16 ——Predicting SR 16 The recursive forecast chain of "".

[0053] Reference Figure 2Given the current time period T, for the forecast time period k=0, the input features of the regression model M0 include the irradiance observation sequence Q(T-1)...Q(Tn) of the previous n time periods and the observation sequence of other meteorological elements Obs(T-1)...Obs(Ti) of the previous i time periods. It also includes the total irradiance value SR(T) of the China-Pacific numerical model at time T and the forecast values ​​HF(T) of other meteorological forecast elements of the China-Pacific numerical model at time T. The model output prediction label Y=SR_pred(T). Figure 2 In the middle, Lead represents the forecast time. For the forecast time k=1, the input features of the regression model M0 still include the irradiance observation sequence and other meteorological element observation sequences of the n times before time T, but only the irradiance forecast value and other meteorological forecast element forecast values ​​for time T+1 are input. The forecast values ​​of other meteorological forecast elements are taken from the Huafeng numerical model HF(T+1), and the irradiance forecast value is taken from the SR_pred(T) output by model M0. The output forecast label Y=SR_pred(T+1).

[0054] Because the forecasts from the Huafeng numerical model inherently contain significant errors, and because it directly uses outdated low-frequency NWP data for modeling, it cannot compensate for the lag in the model's background field, making it difficult to handle extreme weather scenarios such as sudden changes in cloud conditions. This invention combines low-frequency numerical model data with recursive forecasts from a regression model to compensate for the lag.

[0055] Meanwhile, since model M0 has already learned the forecast value of the second-time typhoon during training, to avoid the error propagation affecting the forecasts of subsequent times, model M0 of this invention no longer inputs the forecast value of the second-time typhoon during the T-time period, but only inputs the forecast value of the numerical model forecast elements of the second-time typhoon during the T+1-time period, and uses SR_pred(T) instead of SR(T+1), reducing error propagation and improving prediction accuracy. Subsequent models M... k In addition to inputting the observation sequence, the training only inputs the forecast values ​​of other meteorological forecast elements of the Huafeng numerical model at time T+k, HF(T+k), and the irradiance recursive forecast value SR_pred(T+k-1) at time T+k-1 obtained from the previous model.

[0056] After constructing the recursive forecast architecture, model training is performed, and meteorological element feature selection and hyperparameter localization optimization are completed simultaneously.

[0057] (2) Model Training: Each LightGBM model is defined as a regression task, with the training objective being to minimize the mean squared error (MSE) between the recursive predicted value and the measured irradiance, i.e., the loss function uses MSE. The root mean square error (RMSE) is used as the core training evaluation metric, while the mean absolute error (MAE) and mean absolute percentage error (MAPE) are used as auxiliary evaluation metrics for early stopping judgment and model selection. LightGBM training and validation datasets are constructed, and an early stopping strategy and a maximum number of iterations are set. The early stopping strategy is as follows: during the training of each LightGBM model, the RMSE on the validation set is calculated after each iteration. If the RMSE on the validation set does not decrease after several consecutive iterations (e.g., 50 iterations), training is stopped early, and no further iterations are continued. MAE and MAPE are used as auxiliary evaluation metrics, which can be used for auxiliary observation and judgment during training. For example, if the RMSE value is stable but the MAE decreases, the early stopping can be appropriately relaxed.

[0058] Mean Square Error (MSE): ;in, For recursive forecast values, For the observed values, The total number of samples. MSE measures the average level of the squared forecast error, reflecting the combined magnitude of the error variance (i.e., dispersion) and systematic bias. It is most sensitive to large errors, and its unit is W² / m². 4 .

[0059] Root Mean Square Error (RMSE): RMSE also measures the systematic bias and dispersion of forecast errors, and is more sensitive to large errors. Its unit is W / m².

[0060] Mean Absolute Error (MAE): ; used to measure the average deviation of forecast error, with units of W / m².

[0061] Mean Absolute Percentage Error (MAPE): ;in, The number of valid samples during the day ( ), only during the day (observation value) ) Calculates and measures the average size of the relative error (i.e., the average percentage deviation from the true value), in units of %.

[0062] Furthermore, during training and recursive prediction, the recursive forecast values ​​are labeled for day and night: the solar altitude angle is calculated in real time based on the station's latitude, longitude, and time information. When the solar altitude angle h > 0, the recursive forecast value is retained; when h ≤ 0, the irradiance recursive forecast value is forcibly set to zero. This eliminates the small non-physical positive value outputs that the machine learning model may produce at night and ensures that no non-physical values ​​are introduced into the recursive chain.

[0063] (3) Screening of meteorological element characteristics

[0064] To further improve forecast accuracy, this embodiment further filters the other observational data and other Huafeng numerical model forecast data (forecast data excluding total radiation) obtained in step one. For all candidate variables (including different altitudes of the same variable) of other meteorological forecast elements in the Huafeng numerical model, variables potentially related to irradiance (such as cloud cover, humidity, and wind speed) are initially selected based on radiative transfer and boundary layer physics. Then, through controlled variable experiments: based on the simplest model containing only irradiance and no other variables, candidate variables are added one by one with different lag steps, and the RMSE of the model on the test set is observed to significantly decrease. If the addition of a variable can consistently and significantly reduce the RMSE, and the variable has a physically reasonable explanatory path, it is considered to have "clear physical indicative significance." When RMSEs are similar, MAE and MAPE are used as evaluation indicators. A subset of Huafeng meteorological forecast elements with clear physical indicative significance for radiation forecasts is retained. Similarly, other observational data are also systematically filtered.

[0065] (4) Localized optimization of hyperparameters: Sensitivity tests were conducted on the core hyperparameters (including the number of leaf nodes num_leaves, learning rate, feature sampling ratio feature_fraction, sample sampling ratio bagging_fraction, and sampling frequency bagging_freq). By iterating through the preset value range of each parameter, the optimal combination of hyperparameters was determined with the average RMSE improvement rate compared with the original total radiation forecast of the Chinese wind at each time point as the optimization objective.

[0066] Step 5: Model Forecast Performance Validation and Evaluation

[0067] Based on a time-ordered test set, the model for each forecast time period under the recursive forecast strategy is evaluated step-by-step. The RMSE, MAE, daytime MAPE, and improvement rate compared to the original total radiation forecast for the Chinese winds are calculated for each time period. The improvement rate represents the percentage reduction in error between the original model and this model, objectively reflecting the ultra-short-term forecast gain of this model. .

[0068] For each forecast time period, a combined evaluation chart is generated, including a scatter plot, an error distribution histogram, a time series comparison chart of the complete test set, and a magnified local plot. This visually verifies the model's ability to capture trends in irradiance changes and its forecast accuracy. The indicators from each time period are summarized to verify whether there is an accuracy decay phenomenon under the recursive forecasting strategy.

[0069] Step Six: Operationalizing Ultra-Short-Term Forecasts by Rolling Every 15 Minutes

[0070] (1) Forecast triggering mechanism: In the business operation environment, the rolling forecast process is automatically triggered every 15 minutes to update the latest input data in real time.

[0071] (2) Preparation of forecast input data: For a single forecast mission, the input data includes: ① the forecast data of the numerical model at the latest forecast time for 15 minutes every 0-4 hours in the future; ② the measured irradiance data and other observation element data of the target station at the most recent time.

[0072] (3) Recursive prediction: In the order of the recursive prediction chain, models M0 to M1 are called sequentially. 16 For each model, the total solar radiation (SR) is replaced with the original China wind forecast value by the recursive forecast value from the preceding model. Historical measured irradiance characteristics and other observational features are fixed to the measured values ​​at the corresponding times closest to the forecast start time. Each model independently loads its corresponding model file, and updates the SR recursive forecast value in the recursive chain after completing the prediction.

[0073] (4) Physical constraint processing: The initial predicted values ​​output by the model are processed for physical rationality using the day and night markers of each forecast time. If the day and night marker is 0 (i.e., night), the recursive forecast value of irradiance for that time is forced to be set to 0.

[0074] (5) Forecast results output: The final recursive forecast values ​​of all forecast times are integrated in chronological order to form and output the ultra-short-term solar irradiance rolling forecast sequence for the next 0-4 hours in 15-minute increments.

[0075] (6) Forecast result evaluation: The operational results are statistically summarized at three levels: daily, monthly and overall. Core indicators such as RMSE, MAPE and MAE are calculated, and multi-dimensional visualization verification results such as time series comparison chart, 15-minute MAPE comparison chart and daily average MAPE comparison chart are generated to evaluate the forecast performance and stability of the method of this invention under actual operational rolling conditions.

[0076] The method described in this invention will be verified through specific examples below.

[0077] This embodiment uses the Lüsi surface meteorological station in Jiangsu Province (32.0994°N, 121.5444°E) as the implementation object to conduct ultra-short-term solar irradiance forecasts from 0 to 4 hours, updated every 15 minutes. The specific implementation steps and technical effects of this invention are described in detail below:

[0078] S1, Multi-source data acquisition and preprocessing

[0079] 1) Data acquisition and quality control of measured irradiance at ground stations

[0080] Historical measured surface solar irradiance data from the Lüsi station were retrieved, broken down into 15-minute intervals. The data covers the period from August 1, 2025 to March 31, 2026. Quality control was performed on the irradiance series: the date and time columns were merged into a single "Datetime" column, and data for whole 15-minute intervals (0, 15, 30, and 45 minutes) were retained. Anomaly detection revealed a 150-minute data gap between 04:30 and 07:00 on October 20, 2025, which has been flagged.

[0081] 2) Numerical model forecast data acquisition and spatial matching

[0082] Batch extraction of data from the Huafeng numerical model in NC format compressed package: Valid forecast files were filtered using a dynamic time threshold method. Duplicate forecast data were removed by using adjacent forecast times (such as the next 08:00 or 20:00) as the time threshold, ultimately obtaining model forecast files for 15 minutes from August 8, 2025 to April 11, 2026. Data integrity checks revealed 80 missing data points from 00:00 to 19:45 on November 10, 2025, which have been marked.

[0083] For the Lüsi station, core forecast variables were extracted from the gridded forecast field output by the model. In addition to total radiation (SR), single-layer variables such as total cloud cover (TCC), low cloud cover (LCC), middle cloud cover (MCC), high cloud cover (HCC), cumulative precipitation in the first 15 minutes (PRE_15m), and gusts (GUST) were extracted, along with forecast values ​​for five variables—temperature (TEM), humidity (RHU), wind speed (WS), wind direction (WD), and air pressure (PRS)—at eight altitude levels (10m, 30m, 50m, 70m, 100m, 120m, 150m, and 200m). The 3×3 grid averaging method was used to extract the time series of meteorological elements at the station location, achieving spatial matching between the model gridded field and station observations.

[0084] 3) Data collection of other observation elements at ground stations

[0085] The surface meteorological observation elements of Lüsi Station were obtained every 15 minutes, including air pressure (PRS), sea level pressure (PRS_Sea), air temperature (TEM), dew point temperature (DPT), relative humidity (RHU), water vapor pressure (VAP), instantaneous wind speed (WIN_S_INST), instantaneous wind direction (WIN_D_INST), hourly cumulative precipitation (PRE), and horizontal visibility (VIS_HOR_1MI, VIS_HOR_10MI).

[0086] S2, Data Cleaning and Alignment

[0087] The Huafeng model forecast data, station-measured irradiance data, and other observation data were merged by external join along the time dimension. Systematic quality control was performed on the merged data: outliers of wind direction-related variables exceeding the range [0, 360] were forcibly replaced with NaN; values ​​greater than 9 × 10⁻⁶ were removed from the list of outliers. 5 The missing values ​​were replaced with NaN; linear interpolation was used to fill in missing values ​​in each column. Visual checks were performed on all elements to confirm that the data after quality control was physically sound and free of significant anomalies.

[0088] S3, Construction of Training and Test Sets

[0089] The training set start time was set to 16:00 on August 8, 2025. A sliding window process was applied to the merged data, matching historical observation features for each time point.

[0090] The training and test sets are divided with December 31, 2025 as the dividing line: time series samples on or before December 31, 2025, are set as the training set, and time series samples on or after January 1, 2026, are set as the test set. This division ensures that the test set is completely independent of the training set. The training set covers the seasonal changes from summer through autumn to early winter, while the test set covers the seasonal changes from winter to spring. Both sets contain significant seasonal temperature alternation characteristics, which is beneficial for testing the model's ability to predict different stages of seasonal changes.

[0091] S4 employs a recursive forecasting strategy with time-series independent modeling and localized hyperparameter optimization.

[0092] 1) Reference Figure 2 Establish regression model M k(k=0,1,2…16) and trained independently. In this embodiment, five subsets of meteorological forecasting elements for the Chinese winds with clear physical indication significance for radiation forecasting were selected: Total Cloud Cover (TCC), Low Cloud Cover (LCC), 10m Humidity (RHU_10m), 100m Wind Speed ​​(WS_100m), and 70m Wind Direction (WD_70m). Other ground observation elements were simultaneously screened, and 11 elements were added to the model one by one to test their improvement effect on forecast accuracy. Four elements with better improvement effects were selected: Instantaneous Wind Speed ​​(WIN_S_INST), Relative Humidity (RHU), 1-mile Horizontal Visibility (VIS_HOR_1MI), and 10-mile Horizontal Visibility (VIS_HOR_10MI). Multivariate combination tests were conducted on these four elements, and the results showed that the forecast results were most significantly improved when only Instantaneous Wind Speed ​​(WIN_S_INST) was added. Finally, Instantaneous Wind Speed ​​(WIN_S_INST) was selected as an auxiliary feature. The lag steps of each observation element were optimized through testing. The lag steps of the measured irradiance were selected as 13 steps (n=13), and the lag steps of the instantaneous wind speed were selected as 11 steps (i=11).

[0093] 2) The optimal configuration of the core hyperparameters in this embodiment is: num_leaves=8, learning_rate=0.05, feature_fraction=0.90, bagging_fraction=0.80, bagging_freq=1.

[0094] 3) Independently construct LightGBM training and validation sets for each time interval. Set the model framework to GBDT boosting type, the objective function to a regression function, and RMSE as the core training evaluation metric, with MAE as the auxiliary evaluation metric. Set an early stopping strategy: automatically stop training if the validation set RMSE does not decrease within 50 epochs, with a maximum of 1000 epochs. During training and recursive prediction, perform day / night labeling on the output values ​​of each model.

[0095] S5, Model Forecast Performance Validation and Evaluation

[0096] Figure 3 and Figure 4 The charts show scatter plots, error distribution, and time series combination analysis of the LightGBM recursive forecast values ​​with the Huafeng forecast values ​​and observed values ​​for the next 15 minutes (Lead=1) and the next 4 hours (Lead=16). Figure 3 and Figure 4 (a) in the figure are scatter plots of LightGBM recursive forecast values, original irradiance forecast values ​​of the Chinese wind, and ground-measured irradiance observation values. Figure 3 (b) in the figure is a comparison of the error distribution between the LightGBM recursive forecast and the China Wind forecast. Figure 3 (c) in the figure is a time series comparison of LightGBM recursive forecast values ​​with China wind forecast values ​​and observed values ​​on the complete test set. Figure 3 (d) in the figure shows a comparison of local time series over the most recent two days. At forecast time Lead=1, compared to the original numerical forecast for the Huafeng model, this invention improves RMSE by 30.6% and MAE by 44.9% on the test set; at forecast time Lead=16, RMSE improves by 25.7% and MAE by 37.5%. Overall, the RMSE and MAE indicators remain stable across forecast times under the recursive forecast strategy of this invention, without exhibiting the significant accuracy degradation after the second forecast time common in traditional strategies. The RMSE improvement rate for each forecast lead exceeds 25%, and the overall RMSE decreases from 67.5 W / m² to approximately 44 W / m², effectively solving the problem of systematically low irradiance forecasts in the Huafeng model and fully meeting the operational accuracy requirements for ultra-short-term photovoltaic forecasts.

[0097] S6, operationalizing rolling ultra-short-term forecasts every 15 minutes.

[0098] In a simulated real-world operational environment, with the evaluation period from January 10th to March 31st, 2026, a rolling forecast process was triggered every 15 minutes. The independent models for each time period were sequentially called according to the recursive forecast chain for prediction, and after day / night labeling, a forecast sequence for the next 0-4 hours was output in 15-minute increments. More than 7,000 forecast triggers were conducted during the evaluation period.

[0099] The operational results are statistically summarized at three levels: monthly, daily, and overall, to calculate core indicators such as MAPE, RMSE, and MAE. For selected target forecast periods, multi-dimensional visualization verification results are generated, including time series comparison charts, 15-minute MAPE comparison charts, and daily average MAPE comparison charts. Figure 5 and Figure 6 The charts show the time series of recursive forecasts for the next 15 minutes (Lead=1) and the next 4 hours (Lead=16), combined with the forecast and observed values ​​of the Chinese winds, as well as the 15-minute MAPE comparison and the daily average MAPE comparison. Figure 5 and Figure 6 (a) are time series comparison charts of irradiance, (b) are MAPE comparison charts (daytime), and (c) are daily average MAPE comparison charts.

[0100] The quantitative assessment results during the assessment period are shown in Table 1:

[0101] Table 1 Quantitative assessment results during the assessment period

[0102]

[0103] Note: In Table 1, LGB represents the LightGBM recursive forecast result, and HF represents the Huafeng model forecast result.

[0104] As can be seen from the above results, the method of the present invention is consistently and significantly better than the original forecast of the Chinese wind at all forecast times; the RMSE, MAPE and MAE indices at all forecast times remain stable, and there is no phenomenon of significant decay of forecast accuracy as the lead increases.

[0105] The present invention provides an ultra-short-term solar irradiance rolling forecast system, comprising:

[0106] The measured data acquisition unit is used to acquire historical measured sequences of solar irradiance and historical measured sequences of meteorological observation elements.

[0107] The numerical model forecast data acquisition unit is used to acquire solar irradiance model forecast sequences and meteorological forecast element model forecast sequences based on the Huafeng numerical model.

[0108] The recursive forecast unit is used to independently build a LightGBM regression model for each forecast time, forecast the recursive forecast value of solar irradiance for that forecast time, and integrate the recursive forecast values ​​in chronological order to obtain a solar irradiance forecast sequence.

[0109] For the k-th forecast time, the output of the corresponding LightGBM regression model is the solar irradiance model forecast value SR_pred(T+k) for the k-th forecast time. The input features of the LightGBM regression model include the historical measured sequence of solar irradiance, the historical measured sequence of meteorological observation elements, the model forecast value of meteorological forecast elements for the k-th forecast time, and SR_pred(T+k-1). Among them, the model forecast value of meteorological forecast elements is taken from the model forecast sequence of meteorological forecast elements. When k=0, SR_pred(T+k-1) is replaced by the total radiative radiation SR(T) of the Chinese wind at time T in the model forecast sequence of solar irradiance.

[0110] The electronic device of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded onto the processor, it implements the ultra-short-term solar irradiance rolling forecast method.

[0111] The computer program product of the present invention includes a computer program that, when executed by a processor, implements the ultra-short-term solar irradiance rolling forecast method.

[0112] The computer-readable storage medium may include RAM, ROM, EEPROM, CD-ROM or other optical disc storage devices, magnetic disk storage devices or other magnetic storage devices, flash memory, or any other media that can be used to store program code in the form of instructions or data structures and is accessible by a computer.

[0113] The processor is used to execute a computer program stored in memory to implement the various steps in the methods described in the above embodiments.

Claims

1. A method of ultra-short-term solar irradiance rolling prediction, characterized in that, Includes the following steps: Obtain historical measured sequences of solar irradiance and historical measured sequences of meteorological observation elements; The solar irradiance model forecast sequence and the meteorological forecast element model forecast sequence were obtained based on the Huafeng numerical model. For each forecast time, an independent LightGBM regression model is established to predict the recursive forecast value of solar irradiance for that forecast time. The recursive forecast values ​​are then integrated in chronological order to obtain the solar irradiance forecast sequence. For the k-th forecast time, the output of the corresponding LightGBM regression model is the solar irradiance model forecast value SR_pred(T+k) for the k-th forecast time, where T is the current time. The input features of the LightGBM regression model include the historical measured sequence of solar irradiance, the historical measured sequence of meteorological observation elements, the model forecast value of meteorological forecast elements for the k-th forecast time, and SR_pred(T+k-1). Among them, the model forecast value of meteorological forecast elements is taken from the model forecast sequence of meteorological forecast elements. When k=0, SR_pred(T+k-1) is replaced by the total radiative radiation SR(T) of the Chinese wind at time T in the solar irradiance model forecast sequence.

2. The ultra-short term solar irradiance rolling prediction method according to claim 1, characterized in that, Each 15-minute interval within the next 0 to 4 hours is considered a forecast interval, k=0, 1, 2, ..., 16.

3. The ultra-short term solar irradiance rolling prediction method according to claim 1, characterized in that, Meteorological observation elements include at least one of instantaneous wind speed, instantaneous wind direction, hourly cumulative precipitation, and horizontal visibility.

4. The ultra-short-term solar irradiance rolling forecast method according to claim 1, characterized in that, Meteorological forecast elements include at least one of the following: total cloud cover, low cloud cover, middle cloud cover, high cloud cover, cumulative precipitation, temperature, humidity, wind speed, wind direction, gusts, and air pressure.

5. The ultra-short-term solar irradiance rolling forecast method according to claim 1, characterized in that, The method also includes aligning the acquired historical measured sequences of solar irradiance, historical measured sequences of meteorological observation elements, solar irradiance model forecast sequences, and meteorological forecast element model forecast sequences according to the time dimension, and cleaning the data based on physical rationality.

6. The ultra-short-term solar irradiance rolling forecast method according to claim 1, characterized in that, The loss function of the LightGBM regression model is the mean square error between the recursive predicted value and the measured irradiance.

7. The ultra-short-term solar irradiance rolling forecast method according to claim 1, characterized in that, If the forecast time falls at night, the solar irradiance model forecast value for that forecast time will be forced to be 0.

8. A rolling forecasting system for ultra-short-term solar irradiance, characterized in that, include: The measured data acquisition unit is used to acquire historical measured sequences of solar irradiance and historical measured sequences of meteorological observation elements. The numerical model forecast data acquisition unit is used to acquire solar irradiance model forecast sequences and meteorological forecast element model forecast sequences based on the Huafeng numerical model. The recursive forecast unit is used to independently build a LightGBM regression model for each forecast time, forecast the recursive forecast value of solar irradiance for that forecast time, and integrate the recursive forecast values ​​in chronological order to obtain a solar irradiance forecast sequence. For the k-th forecast time, the output of the corresponding LightGBM regression model is the solar irradiance model forecast value SR_pred(T+k) for the k-th forecast time. The input features of the LightGBM regression model include the historical measured sequence of solar irradiance, the historical measured sequence of meteorological observation elements, the model forecast value of meteorological forecast elements for the k-th forecast time, and SR_pred(T+k-1). Among them, the model forecast value of meteorological forecast elements is taken from the model forecast sequence of meteorological forecast elements. When k=0, SR_pred(T+k-1) is replaced by the total radiative radiation SR(T) of the Chinese wind at time T in the model forecast sequence of solar irradiance.

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 computer program is loaded into the processor, it implements the ultra-short-term solar irradiance rolling forecast method according to any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the ultra-short-term solar irradiance rolling forecast method according to any one of claims 1-7.

Images