Data-driven photovoltaic power generation prediction method and system
By decomposing meteorological factors of photovoltaic power plants at multiple time scales and fusing them with multi-head attention, high-precision photovoltaic power prediction results are generated. This solves the problems of limited accuracy in photovoltaic power prediction and insufficient reliability of probability prediction intervals in existing technologies, and enables efficient prediction in scenarios of sudden weather changes.
Patent Information
- Application Number
- CN202611446811.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-17
- Publication Date
- 2026-11-10
AI Technical Summary
Existing photovoltaic power prediction methods fail to effectively distinguish the differentiated impacts of different meteorological factors at different time scales, resulting in limited prediction accuracy. In particular, the prediction bias increases significantly under sudden weather changes, and the reliability of the probability prediction interval is difficult to guarantee.
Employing multi-timescale decomposition and multi-head attention fusion techniques, this study decomposes meteorological factors into short-term fluctuations, intraday trends, and long-term trends. Combining temporal feature extraction and spatial distribution features, the data are input into a quantile regression ensemble prediction model to generate probability prediction intervals at different confidence levels. The model performance is then optimized through online calibration and multi-dimensional evaluation.
It improves the accuracy of photovoltaic power prediction and the reliability of the probability prediction range, adapts to dynamic weather changes, and enhances the effectiveness of feature extraction and the accuracy of prediction results.
Smart Images

Figure CN123070418A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of photovoltaic power generation prediction technology, and in particular to a data-driven photovoltaic power generation prediction method and system. Background Technology
[0002] Photovoltaic (PV) power output is comprehensively affected by various meteorological factors such as irradiance, temperature, and cloud cover. Accurate prediction of PV power is crucial for grid dispatch. Existing PV power prediction methods mostly employ deterministic point prediction approaches, treating all meteorological factors equally and failing to differentiate the varying impacts of different meteorological factors on PV output at different time scales. In terms of probabilistic prediction, existing methods typically make pre-defined assumptions about the error distribution, making it difficult to adapt to the non-Gaussian distribution characteristics of PV power under different weather conditions. These problems limit the accuracy of PV power prediction, especially under scenarios of sudden weather changes where prediction bias increases significantly, and the reliability of the probabilistic prediction interval is difficult to guarantee. Summary of the Invention
[0003] The purpose of this invention is to provide a data-driven photovoltaic power generation prediction method and system to solve the above-mentioned problems.
[0004] To achieve the above objectives, the present invention adopts the following technical solution: Data-driven photovoltaic power generation prediction methods include: Collect multi-source historical data of photovoltaic power plants, decompose each meteorological factor into a multi-time scale, and separate each meteorological factor into short-term fluctuation component, intraday trend component and long-term trend component. The temporal dimension features of each time scale component are extracted through the corresponding temporal feature extraction channels, and the spatial distribution features of meteorological factors are extracted. The temporal dimension features and spatial distribution features are fused through multi-head attention to obtain a spatiotemporal fusion feature vector. The spatiotemporal fusion feature vector is input into the quantile regression integrated prediction model, which outputs photovoltaic power prediction values at multiple target quantiles. Based on the prediction values of each quantile, probability prediction intervals with different confidence levels are generated.
[0005] Preferably, the multi-timescale decomposition is achieved using empirical mode decomposition, which separates the original sequence of meteorological factors into multiple intrinsic mode function components, and classifies them according to the characteristic frequency of each intrinsic mode function component: intrinsic mode function components with characteristic frequencies higher than a first frequency threshold are classified into short-term fluctuation components, intrinsic mode function components with characteristic frequencies between the first and second frequency thresholds are classified into intraday trend components, and intrinsic mode function components with characteristic frequencies lower than the second frequency threshold are merged with the residual trend term and classified into long-term trend components.
[0006] Preferably, the time-series feature extraction channel is a parallel three-channel structure. The short-term channel uses a gated recurrent unit to process the short-term fluctuation component sequence of each meteorological factor and outputs a short-term time-series feature vector. The intraday channel uses a Transformer encoder to process the intraday trend component sequence of each meteorological factor and outputs an intraday time-series feature vector. The long-term channel uses a temporal convolutional network to process the long-term trend component sequence of each meteorological factor and outputs a long-term time-series feature vector.
[0007] Preferably, the spatial distribution features are extracted by dividing the grid around the geographical coordinates of the photovoltaic power station, mapping the meteorological data of each meteorological observation point within different distance ranges around the power station to the corresponding grid nodes, forming a spatial meteorological feature matrix, wherein the rows of the spatial meteorological feature matrix correspond to the grid nodes and the columns correspond to each meteorological factor.
[0008] Preferably, the process of generating the spatiotemporal fusion feature vector is as follows: the short-term time series feature vector, intraday time series feature vector, and long-term time series feature vector are concatenated with the flattened vector of the spatial meteorological feature matrix to form a concatenated feature vector. The concatenated feature vector is used as the input sequence of multi-head attention. Through linear transformation of the query matrix, key matrix, and value matrix and scaling dot product attention operation, attention weights are adaptively assigned to meteorological factors at different spatiotemporal locations. The output of the multi-head attention is linearly projected to obtain the spatiotemporal fusion feature vector.
[0009] Preferably, the quantile regression ensemble prediction model uses a gradient boosting decision tree as the base learner, and configures multiple quantile objective functions for parallel training. The loss function corresponding to each quantile is: ;
[0010] in, For the first The actual power value of each sample For the first The sample at the quantile The predicted value below, For the sample size, Let quantile test function, when hour ,when hour , This represents the residual between the actual power value and the predicted value.
[0011] Preferably, the method further includes: real-time acquisition of the most recent data. The actual power values within each time step are calculated, and the deviation sequence between the actual power values and the predicted values of each quantile is determined. Based on the deviation sequence, the... The mean over time steps and standard deviation Calculate the first Correction for each quantile:
[0012] ;
[0013] in, For the first The proportionality coefficient corresponding to each quantile The preset rolling window length is used; the correction value is superimposed on each quantile prediction value to obtain the calibrated quantile prediction value, and the calibrated probability prediction interval is regenerated.
[0014] Preferably, the update rate of the correction quantity adapts to the weather conditions: when the amplitude of the short-term fluctuation component does not exceed the amplitude threshold within a preset time period, it is determined to be a clear and stable weather condition, and the correction quantity is updated with a small learning rate; when the amplitude of the short-term fluctuation component exceeds the amplitude threshold within a preset time period, it is determined to be a rapidly changing weather condition, and the correction quantity is updated with a large learning rate.
[0015] Preferably, the method further includes: calculating multi-dimensional evaluation indicators, including point prediction indicators, probability prediction indicators, and comprehensive scoring indicators; automatically calculating the values of each evaluation indicator daily and comparing them with preset warning lines; when the point prediction indicator exceeds the corresponding warning line for a preset number of consecutive days, or the coverage rate of the probability prediction interval deviates from its nominal confidence level by more than a preset deviation range, merging the newly accumulated data increment into the training set and re-executing the model training process; and selecting the model with better comprehensive scoring indicators for deployment through comparative verification.
[0016] A data-driven photovoltaic power generation prediction system includes:
[0017] The data decomposition module is used to collect multi-source historical data of photovoltaic power plants, decompose various meteorological factors at multiple time scales, and obtain the short-term fluctuation components, intraday trend components and long-term trend components of each meteorological factor.
[0018] The spatiotemporal feature extraction module is used to extract the temporal dimension features of each time scale component through the corresponding temporal feature extraction channel, extract the spatial distribution features of meteorological factors, and fuse the temporal dimension features and spatial distribution features through multi-head attention to obtain a spatiotemporal fusion feature vector.
[0019] The probability prediction module is used to input the spatiotemporal fusion feature vector into the quantile regression integrated prediction model, output photovoltaic power prediction values at multiple target quantiles, and generate probability prediction intervals with different confidence levels based on the prediction values of each quantile.
[0020] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:
[0021] 1. This invention decomposes meteorological factors into multiple time scales, and sends the short-term fluctuation component, intraday trend component and long-term trend component into matching time-series feature extraction channels respectively. Since the components at different time scales have fundamentally different effects on photovoltaic output, the separation processing allows each channel to focus on the feature patterns at the corresponding time scale, overcoming the problem of mutual interference between different frequency information, and helping to improve the characterization accuracy of the influence of meteorological factors on photovoltaic output.
[0022] 2. This invention integrates temporal and spatial distribution features based on multi-head attention, and adaptively assigns differentiated weights to meteorological factors with different distance ranges and time scales. This overcomes the limitation of fixed weights in adapting to dynamic meteorological changes and helps improve the effectiveness of feature extraction. Attached Figure Description
[0024] Further details, features, and advantages of this application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which:
[0025] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0027] Several embodiments of this application will now be described in more detail with reference to the accompanying drawings to enable those skilled in the art to implement this application. This application may be embodied in many different forms and for various purposes and should not be limited to the embodiments set forth herein. These embodiments are provided to make this application thorough and complete, and to fully convey the scope of this application to those skilled in the art. The embodiments described do not limit this application.
[0028] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It will be further understood that terms such as those defined in commonly used dictionaries shall be interpreted as having a meaning consistent with their meaning in the relevant field and / or the context of this specification, and shall not be interpreted in an idealized or overly formal sense unless expressly defined herein.
[0029] Example 1
[0030] Its specific implementation method is combined with the appendix Figure 1 Please provide a detailed explanation.
[0031] In this embodiment, it includes:
[0032] In the field of photovoltaic (PV) power generation forecasting, existing methods mostly employ deterministic point prediction, treating all meteorological factors equally and failing to distinguish the differentiated impacts of meteorological factors on PV output at different time scales. Furthermore, existing methods struggle to effectively quantify forecast uncertainty, failing to provide reliable probabilistic prediction intervals for grid dispatch. These issues limit the accuracy of PV power forecasting, especially under scenarios of sudden weather changes, where forecast bias increases significantly, and the reliability of probabilistic prediction intervals is difficult to guarantee.
[0033] A photovoltaic power generation probability prediction method according to an embodiment of the present invention includes the following steps:
[0034] Step 1: Collect multi-source historical data from the photovoltaic power station, decompose meteorological factors into multiple time scales, and obtain the components of each meteorological factor at different time scales.
[0035] Historical power generation data of the photovoltaic power plant and corresponding multi-source meteorological data were collected. The multi-source meteorological data included numerical weather prediction data, satellite cloud imagery data, and measured data from ground meteorological stations. The numerical weather prediction data included total radiation, temperature, humidity, wind speed, and cloud cover. Before inputting the data into the model, the historical power generation data and each meteorological factor data were Z-score standardized to eliminate the impact of differences in the dimensions of different physical quantities on subsequent calculations.
[0036] Each meteorological factor was decomposed into three time-scale components: short-term fluctuation component, intraday trend component, and long-term trend component. The short-term fluctuation component, ranging from minutes to hours, reflects the impact of rapid cloud movement on irradiance. The intraday trend component, ranging from hours to days, reflects the impact of changes in solar altitude angle on power output. The long-term trend component, ranging from days to months, reflects the impact of seasonal variations and meteorological system evolution.
[0037] It should be noted that the above multi-timescale decomposition is implemented using empirical mode decomposition or wavelet decomposition. The input is the original time series of each meteorological factor, and the output is the corresponding three time-scale components. Taking empirical mode decomposition as an example, the original sequence of meteorological factors is adaptively separated into multiple intrinsic mode function (IMF) components. These IMF components are then classified according to their characteristic frequencies: IMF components with characteristic frequencies higher than a first frequency threshold are classified as short-term fluctuation components; IMF components with characteristic frequencies between the first and second frequency thresholds are classified as intraday trend components; and IMF components with characteristic frequencies lower than the second frequency threshold are merged with the residual trend term and classified as long-term trend components. The first frequency threshold corresponds to a period of approximately 1 hour, and the second frequency threshold corresponds to a period of approximately 24 hours. Both thresholds can be adjusted according to the spectral distribution of the actual power plant data.
[0038] Step 2: Extract spatiotemporal features of meteorological factors based on multi-head attention to obtain spatiotemporal fusion feature vectors.
[0039] Based on short-term fluctuation components, intraday trend components, and long-term trend components, time-dimensional features are extracted through corresponding time-series feature extraction channels. Simultaneously, based on the geographical grid division of the photovoltaic power station area, spatial distribution features of meteorological factors are extracted. Multi-head attention is used to fuse the time-dimensional features and spatial distribution features to obtain a spatiotemporal fused feature vector.
[0040] It should be noted that the above-mentioned time-series feature extraction channels are parallel three-channel structures, with each channel processing the components corresponding to its time scale. The short-term channel uses a gated recurrent unit, whose input is the short-term fluctuation component sequence of each meteorological factor, and whose output is a short-term time-series feature vector; the intraday channel uses a Transformer encoder, whose input is the intraday trend component sequence of each meteorological factor, and whose output is an intraday time-series feature vector; the long-term channel uses a temporal convolutional network, whose input is the long-term trend component sequence of each meteorological factor, and whose output is a long-term time-series feature vector.
[0041] It should be noted that the above spatial distribution features are extracted as follows: A grid is created centered on the geographical coordinates of the photovoltaic power station, and meteorological data from various meteorological observation points within different distances around the power station are mapped to corresponding grid nodes, forming a spatial meteorological feature matrix. The rows of the spatial meteorological feature matrix correspond to grid nodes, and the columns correspond to various meteorological factors.
[0042] It should be noted that the generation process of the aforementioned spatiotemporal fusion feature vector is as follows: the short-term time-series feature vector, intraday time-series feature vector, and long-term time-series feature vector are concatenated with the flattened vector of the spatial meteorological feature matrix to form a concatenated feature vector. This concatenated feature vector is used as the input sequence for multi-head attention. Through linear transformations of the query matrix, key matrix, and value matrix, as well as scaling dot product attention operations, attention weights are adaptively assigned to meteorological factors at different spatiotemporal locations. Spatially, meteorological factors within a short distance receive higher attention weights, while those within a long distance receive moderate attention weights. The output of the multi-head attention is then linearly projected to obtain the spatiotemporal fusion feature vector.
[0043] Step 3: Input the spatiotemporal fusion feature vector into the quantile regression ensemble prediction model, output the photovoltaic power prediction values at multiple quantiles, and generate the probability prediction interval.
[0044] The spatiotemporal fusion feature vector is input into the quantile regression ensemble prediction model, which outputs photovoltaic power prediction values at multiple target quantiles. Based on each quantile prediction value, probability prediction intervals with different confidence levels are generated, and the width of the prediction interval is calculated as a quantitative indicator of prediction uncertainty. Since the output of the quantile regression ensemble prediction model is a power prediction value standardized by Z-score, each quantile prediction value needs to be destandardized to restore it to the prediction result in the actual power dimension before it is used to generate the probability prediction interval.
[0045] It should be noted that the above quantile regression ensemble prediction model uses a gradient boosting decision tree as its base learner, and its input is a spatiotemporal fusion feature vector. Five quantile objective functions (5%, 25%, 50%, 75%, and 95%) are configured for parallel training. Let the ... The loss function corresponding to each quantile is the quantile loss. Its definition is:
[0046] ;
[0047] in, For the first The actual power value of each sample For the first The sample at the quantile The predicted value below, For the sample size, For sample index, The quantile test function is defined as follows:
[0048] ;
[0049] in, This represents the residual between the actual and predicted power values. The objective function for each quantile independently optimizes the corresponding quantile loss, outputting the predicted photovoltaic power value at each quantile.
[0050] It should be noted that the probability prediction intervals are generated as follows: a 50% confidence level prediction interval is formed by the 25th and 75th quantile predicted values, and a 90% confidence level prediction interval is formed by the 5th and 95th quantile predicted values. The 50th quantile predicted value is used as the median prediction value. The width of the prediction interval is the difference between the corresponding upper and lower quantile predicted values.
[0051] In this embodiment of the application, in order to eliminate the systematic bias of the quantile regression ensemble prediction model caused by factors such as sudden weather changes, equipment degradation, and dust accumulation during long-term operation, the following steps are also included in step 3:
[0052] Step 4: Perform online calibration of the prediction results based on the dynamic quantile bias to obtain the calibrated probability prediction interval.
[0053] Real-time collection of recent data The actual power values within each time step are calculated, and the deviation sequence between the actual power values and the predicted values of each quantile output in step 3 is calculated. The preset rolling window length is used. Based on the rolling statistics of the bias sequence, the correction amount for each quantile prediction value is dynamically calculated. The correction amount is then added to each quantile prediction value to obtain the calibrated quantile prediction value, and the calibrated probability prediction interval is regenerated. Here, the bias sequence is calculated in the actual power dimension after destandardization, and the dimensions of the correction amount are consistent with those of each quantile prediction value.
[0054] It should be noted that the rolling statistics mentioned above include the deviation series in... The mean over time steps and standard deviation For the first Each quantile, its correction amount The calculation method is as follows:
[0055] ;
[0056] in, For the biased sequence in The mean over each time step For the biased sequence in Standard deviation within each time step For the first The proportionality coefficient corresponding to each quantile This is the preset scroll window length. For the 50th percentile, Setting it to 0 degenerates the correction amount into the mean deviation. For quantiles below 50%, Taking a negative value shifts the correction downwards; for quantiles above 50%, Take a positive value to shift the correction upwards. The calibrated [number]th [unit / item]. The quantile prediction value is the original prediction value minus the correction. sum.
[0057] It should be noted that the update rate of the above correction values adaptively changes with weather conditions. Weather condition indicators are defined as follows: when the amplitude of the short-term fluctuation component does not exceed the amplitude threshold within a preset time period, it is determined to be a stable weather condition, and the correction value is updated slowly with a small learning rate; when the amplitude of the short-term fluctuation component exceeds the amplitude threshold within a preset time period, it is determined to be a rapidly changing weather condition, and the correction value is updated rapidly with a large learning rate. The amplitude threshold is determined based on the amplitude distribution of the short-term fluctuation component during stable and rapidly changing weather conditions in historical data. Both the small and large learning rates are preset parameters and can be adjusted based on actual power plant operating data.
[0058] In this embodiment of the application, in order to achieve long-term adaptive updating of the quantile regression ensemble prediction model and avoid the continuous degradation of the performance of the quantile regression ensemble prediction model as the operating environment changes, the following steps are also included in step 4:
[0059] Step 5: Evaluate the prediction performance from multiple dimensions, trigger automatic model retraining based on the evaluation results, and select the optimal model for deployment.
[0060] Calculate multi-dimensional evaluation metrics, including point prediction metrics, probabilistic prediction metrics, and a comprehensive scoring metric. Point prediction metrics include mean absolute percentage error and root mean square error. Probabilistic prediction metrics include prediction interval coverage and average prediction interval width. The comprehensive scoring metric is based on the coverage width criterion.
[0061] The system automatically calculates the values of each evaluation indicator daily and compares them with preset warning lines. When the retraining trigger condition is met, the newly accumulated data increment is merged into the training set, and the model training process of steps 2 to 3 is re-executed to update the weight parameters of the quantile regression ensemble prediction model.
[0062] It should be noted that the above retraining trigger condition is that either of the following conditions is met: the point prediction indicator exceeds the corresponding warning line for a consecutive preset number of days; or the coverage rate of the probability prediction interval deviates from its nominal confidence level by more than a preset deviation range. For example, if the nominal coverage rate of the 50% confidence interval is 50%, and its actual coverage rate is continuously lower than 40% or continuously higher than 60%, it is determined to be a coverage deviation, and retraining is triggered.
[0063] It should be noted that after retraining, the new model and the original model are compared and validated using A / B testing. Specifically, the data from a recent period is divided into a validation set, and both the new model and the original model are used to make predictions on the validation set, calculating various evaluation metrics. The model with the better overall score is selected for deployment, replacing the currently running model.
[0064] Example 2
[0065] A photovoltaic power station (code name: GF-NW01) is located in a desert region of a northwestern province, with an installed capacity of 100 MW. Situated in the Gobi Desert, the station experiences frequent sandstorms in spring and autumn, and active convective clouds in the afternoons of summer, leading to drastic short-term fluctuations in irradiance. The power station is connected to the provincial power grid dispatch system, which requires 15-minute probability predictions of photovoltaic power output, along with prediction intervals at 50% and 90% confidence levels, to support the preparation of day-ahead and intraday dispatch plans. The power station has accumulated three years of historical operating data from years A to C, including ground meteorological station measurements, numerical weather prediction results, and satellite cloud imagery.
[0066] The data decomposition module aggregates multi-source historical data from the GF-NW01 power plant. The raw data includes five meteorological factors: total radiation, temperature, humidity, wind speed, and cloud cover, along with measured power generation for the corresponding time periods, with a time resolution of 15 minutes. After Z-score standardization, empirical mode decomposition is applied to the total radiation sequence, adaptively separating several intrinsic mode function components. Components with characteristic frequencies above the first frequency threshold (corresponding to a period of approximately 1 hour) are classified as short-term fluctuation components, reflecting abrupt changes in irradiance caused by cloud cover; components with characteristic frequencies between the first and second frequency thresholds (corresponding to a period of approximately 24 hours) are classified as intra-diurnal trend components, reflecting the diurnal variation of solar altitude angle; components below the second frequency threshold are merged with the residual trend term and classified as long-term trend components, reflecting seasonal evolution and the gradual influence of the meteorological system. The table below shows the decomposition results (standardized dimensions) of the total radiation factor on a typical day in year C at three time scales.
[0067] Table 1. Multi-timescale decomposition results of total radiation factor (standardized dimensions)
[0068]
[0069] The spatiotemporal feature extraction module takes the three time-scale components output from step 1 as input and starts three parallel temporal feature extraction channels. The short-term channel uses a gated recurrent unit to process the short-term fluctuation component sequences of factors such as total radiation and cloud cover, capturing the rapid changes in cloud cover characteristics; the intraday channel uses a Transformer encoder to process the intraday trend component sequences of each factor, capturing the power rise and fall patterns driven by the solar altitude angle; and the long-term channel uses a temporal convolutional network to process the long-term trend component sequences of each factor, extracting seasonal background information.
[0070] At the same time, the grid was divided with the geographical coordinates of the GF-NW01 power station as the center, and the meteorological data of the surrounding meteorological observation points (code: OBS-01 to OBS-05) were mapped to the corresponding grid nodes to form a spatial meteorological feature matrix. The rows of the matrix correspond to the grid nodes, and the columns correspond to the meteorological factors.
[0071] Table 2 Spatial meteorological characteristic matrix (12:00 on a typical day in year C, standardized dimensions)
[0072]
[0073] The temporal feature vectors output from the three channels are concatenated with the flattened vector of the spatial meteorological feature matrix and then fed into the multi-head attention computation. In the spatial dimension, OBS-01 (5 km from the power station) receives the highest attention weight, while OBS-05 (78 km from the power station) receives a relatively lower weight, achieving differentiated fusion of near- and far-distance meteorological factors. The multi-head attention output is linearly projected to obtain a spatiotemporal fusion feature vector, which is then used by the subsequent prediction module.
[0074] The probability prediction module inputs the spatiotemporal fusion feature vector into the quantile regression ensemble prediction model, outputting the corresponding standardized power prediction values under the objective functions of five quantiles: 5%, 25%, 50%, 75%, and 95%. Training for each quantile uses quantile loss. To optimize the objective, among which , For the first The actual power value of each sample For the first The sample at the quantile The predicted value below, For the sample size, This is the quantile test function. Each quantile prediction value is denormalized to its actual power dimension (megawatts) and then combined and output according to the interval generation rules. The table below shows the quantile prediction results and corresponding prediction intervals for a typical day in year C, covering a portion of the time period.
[0075] Table 3 Quantile Prediction Results and Probability Prediction Intervals (Inverse Standardization, Unit: MW)
[0076]
[0077] The 50% confidence interval is composed of the 25th and 75th percentile predicted values, while the 90% confidence interval is composed of the 5th and 95th percentile predicted values. During the 14:00 period, due to active afternoon convective clouds, the short-term fluctuation component experienced larger amplitudes, resulting in significantly wider confidence intervals for both periods compared to the morning period, reflecting a reasonable quantification of forecast uncertainty.
[0078] The online calibration module uses a rolling window length. For 8 time steps (i.e., 2 hours), the actual power values of the GF-NW01 power plant for the most recent 2 hours are collected in real time. The deviation sequence of the predicted values for each quantile is calculated, and the mean of the deviation sequence is calculated. and standard deviation For the first Quantiles, correction amount ,in The 50th percentile is set to 0, the 25th and 5th percentiles are set to negative values, and the 75th and 95th percentiles are set to positive values.
[0079] Taking a sandstorm-affected day in year C as an example, the measured power before noon was consistently lower than the 50th percentile predicted value, with a deviation from the mean. A negative value indicates a systematic overestimation. At this point, the amplitude of the short-term fluctuation component exceeds the amplitude threshold, and the system determines this as a rapidly changing weather condition, updating the correction value quickly with a large learning rate. The statistics for the period at 12:00 on that day are as follows: MW, MW, the correction values for each quantile are based on the formula Calculations. The table below shows a comparison of the predicted values for each quantile before and after calibration.
[0080] Table 4 Comparison of quantile predictions before and after online calibration (unit: MW)
[0081]
[0082] After calibration, the 50% confidence interval was adjusted from (74.5, 92.0) to (68.1, 89.3), with the interval shifting downwards and significantly improving the agreement with the measured power.
[0083] The evaluation and retraining module automatically calculates multi-dimensional evaluation indicators for the GF-NW01 power plant daily, including the mean absolute percentage error and root mean square error of point predictions, the interval coverage and average interval width of probabilistic predictions, and the coverage width criterion score for the comprehensive evaluation. The table below shows the evaluation results and retraining triggers for four consecutive weeks in Year C.
[0084] Table 5 Weekly Summary of Multi-Dimensional Evaluation Indicators and Retraining Trigger Records
[0085]
[0086] In week 4, the 50% interval coverage rate remained below 40% for several consecutive weeks, meeting the retraining trigger condition. The system merged the newly accumulated data increment into the training set and re-executed the training process from steps 2 to 3, generating a new version model (codename: GF-NW01-v2). Subsequently, A / B testing was conducted, using recent data as the validation set, to calculate the coverage width criterion scores for both the old and new models. The coverage width criterion score of GF-NW01-v2 was superior to that of the original model, and the system deployed it online, replacing the currently running model.
[0087] The data flow throughout the entire implementation process has a complete logical chain. The three time-scale components output from step 1 (Table 1) serve as the input to step 2, and together with the spatial meteorological feature matrix (Table 2), they are fused through multi-head attention to generate a spatiotemporal fusion feature vector. The spatiotemporal fusion feature vector from step 2 drives step 3 to output five quantile predicted values and probability prediction intervals (Table 3). The deviation sequence between the predicted values and the measured power from step 3 enters the online calibration in step 4. The correction amount is adaptively adjusted according to the weather conditions, and the comparison before and after calibration (Table 4) reflects the real-time elimination of systematic deviations. Step 5 continuously monitors the comprehensive prediction effect from steps 3 to 4 (Table 5), and initiates retraining when the coverage deviates from the trigger condition to ensure that the model maintains stable prediction performance under the long-term operating environment of the GF-NW01 power plant.
[0088] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.
[0089] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.
[0090] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A data-driven photovoltaic power generation prediction method, characterized in that, include: Collect multi-source historical data of photovoltaic power plants, decompose each meteorological factor into a multi-time scale, and separate each meteorological factor into short-term fluctuation component, intraday trend component and long-term trend component; The temporal dimension features of each time scale component are extracted through the corresponding temporal feature extraction channels, and the spatial distribution features of meteorological factors are extracted. The temporal dimension features and spatial distribution features are fused through multi-head attention to obtain a spatiotemporal fusion feature vector. The spatiotemporal fusion feature vector is input into the quantile regression integrated prediction model, which outputs photovoltaic power prediction values at multiple target quantiles. Based on the prediction values of each quantile, probability prediction intervals with different confidence levels are generated.
2. The data-driven photovoltaic power generation prediction method according to claim 1, characterized in that, Multi-timescale decomposition is achieved using empirical mode decomposition, which separates the original sequence of meteorological factors into multiple intrinsic mode function components. The components are then classified according to their characteristic frequencies: intrinsic mode function components with characteristic frequencies higher than the first frequency threshold are classified into short-term fluctuation components; intrinsic mode function components with characteristic frequencies between the first and second frequency thresholds are classified into intraday trend components; and intrinsic mode function components with characteristic frequencies lower than the second frequency threshold are merged with the residual trend term and classified into long-term trend components.
3. The data-driven photovoltaic power generation prediction method according to claim 1, characterized in that, The temporal feature extraction channel is a parallel three-channel structure. The short-term channel uses a gated recurrent unit to process the short-term fluctuation component sequence of each meteorological factor and outputs a short-term temporal feature vector. The intraday channel uses a Transformer encoder to process the intraday trend component sequence of each meteorological factor and outputs an intraday temporal feature vector. The long-term channel uses a temporal convolutional network to process the long-term trend component sequence of each meteorological factor and outputs a long-term temporal feature vector.
4. The data-driven photovoltaic power generation prediction method according to claim 3, characterized in that, The spatial distribution features are extracted as follows: grid division is carried out with the geographical coordinates of the photovoltaic power station as the center, and the meteorological data of each meteorological observation point within different distance ranges around the power station are mapped to the corresponding grid nodes to form a spatial meteorological feature matrix. The rows of the spatial meteorological feature matrix correspond to the grid nodes, and the columns correspond to each meteorological factor.
5. The data-driven photovoltaic power generation prediction method according to claim 4, characterized in that, The process of generating the spatiotemporal fusion feature vector is as follows: the short-term time series feature vector, intraday time series feature vector, and long-term time series feature vector are concatenated with the flattened vector of the spatial meteorological feature matrix to form a concatenated feature vector. The concatenated feature vector is used as the input sequence of multi-head attention. Through linear transformation of the query matrix, key matrix, and value matrix and scaling dot product attention operation, attention weights are adaptively assigned to meteorological factors at different spatiotemporal locations. The output of the multi-head attention is linearly projected to obtain the spatiotemporal fusion feature vector.
6. The data-driven photovoltaic power generation prediction method according to claim 1, characterized in that, The quantile regression ensemble prediction model uses a gradient boosting decision tree as the base learner, and is trained in parallel with multiple quantile objective functions configured separately. The loss function corresponding to each quantile is: ; in, For the first The actual power value of each sample For the first The sample at the quantile The predicted value below, For the sample size, Let quantile test function, when hour ,when hour , This represents the residual between the actual power value and the predicted power value.
7. The data-driven photovoltaic power generation prediction method according to claim 1, characterized in that, Also includes: Real-time collection of recent data The actual power values within each time step are calculated, and the deviation sequence between the actual power values and the predicted values of each quantile is determined. Based on the deviation sequence, the... The mean over time steps and standard deviation Calculate the first Correction for each quantile: ; in, For the first The proportionality coefficient corresponding to each quantile The preset rolling window length is used; the correction value is superimposed on each quantile prediction value to obtain the calibrated quantile prediction value, and the calibrated probability prediction interval is regenerated.
8. The data-driven photovoltaic power generation prediction method according to claim 1, characterized in that, The update rate of the correction quantity adapts to the weather conditions: when the amplitude of the short-term fluctuation component does not exceed the amplitude threshold within a preset time period, it is determined to be a clear and stable weather condition, and the correction quantity is updated with a small learning rate; when the amplitude of the short-term fluctuation component exceeds the amplitude threshold within a preset time period, it is determined to be a rapidly changing weather condition, and the correction quantity is updated with a large learning rate.
9. The data-driven photovoltaic power generation prediction method according to claim 7, characterized in that, Also includes: Calculate multi-dimensional evaluation metrics, including point prediction metrics, probability prediction metrics, and comprehensive scoring metrics. Calculate the values of each evaluation metric automatically daily and compare them with preset warning lines. When the point prediction metric exceeds the corresponding warning line for a preset number of consecutive days, or the probability prediction interval coverage deviates from its nominal confidence level by a preset deviation range, merge the newly accumulated data increment into the training set and re-execute the model training process. Through comparative verification, select the model with the better comprehensive scoring metrics for deployment.
10. A data-driven photovoltaic power generation prediction system. The data-driven photovoltaic power generation prediction method according to any one of claims 1-9 is characterized in that, include: The data decomposition module is used to collect multi-source historical data of photovoltaic power plants, decompose various meteorological factors at multiple time scales, and obtain the short-term fluctuation components, intraday trend components and long-term trend components of each meteorological factor. The spatiotemporal feature extraction module is used to extract the temporal dimension features of each time scale component through the corresponding temporal feature extraction channel, extract the spatial distribution features of meteorological factors, and fuse the temporal dimension features and spatial distribution features through multi-head attention to obtain a spatiotemporal fusion feature vector. The probability prediction module is used to input the spatiotemporal fusion feature vector into the quantile regression integrated prediction model, output photovoltaic power prediction values at multiple target quantiles, and generate probability prediction intervals with different confidence levels based on the prediction values of each quantile.
