Short-term photovoltaic power prediction method and device based on causal homogeneous large sample pool and medium

Abstract

The application provides a short-term photovoltaic power prediction method and device based on causal homogeneity large sample pool and medium, relating to the technical field of new energy power generation prediction. The method comprises the following steps: obtaining historical power and numerical weather prediction data of a target photovoltaic power station and a reference photovoltaic power station, and constructing a cross-station basic large sample pool after capacity equivalent conversion; dividing the climbing period, the stable period and the attenuation period according to the daily output law, calculating the normalized comprehensive causal importance score NCCIS of different stages, and purifying the sample pool in a causal homogeneity; matching similar numerical weather prediction sequences in the purified sample pool, extracting the most similar historical power segment to form a historical response prior, and inputting the prediction model to obtain a short-term photovoltaic power prediction result. The application can reduce non-causal similar sample interference, improve the prediction stability, migration and practicability in complex weather and cross-station reuse scenarios.

Description

Technical Field

[0001] This application relates to the field of new energy power generation prediction technology, and in particular to a short-term photovoltaic power prediction method, device and medium based on a large sample pool of causal homogeneity. Background Technology

[0002] Currently, the installed capacity of new energy sources continues to increase rapidly, and the global energy structure transformation is accelerating. Photovoltaic power generation is significantly affected by solar radiation, cloud cover, temperature, humidity, and local meteorological conditions, exhibiting randomness, intermittency, and volatility. With the continuous expansion of high-proportion grid-connected photovoltaic power, the impact of photovoltaic output fluctuations on the safe and stable operation of the power system, reserve capacity allocation, and new energy consumption is becoming increasingly prominent. Therefore, accurately conducting short-term power forecasting for photovoltaic power plants is of great significance in ensuring regional power balance.

[0003] Short-term photovoltaic (PV) power forecasting typically refers to power predictions for the next few hours to the next day from the forecast time, with a commonly used time resolution of 15 minutes. The significance of short-term PV power forecasting lies in providing a reference for day-ahead planning, dispatching operations, reserve capacity preparation, power deviation assessment, and renewable energy consumption decisions.

[0004] While there are numerous existing studies on photovoltaic power prediction, several shortcomings remain: First, most methods rely primarily on historical power data from a single station and numerical weather prediction data for modeling, failing to fully utilize the common physical responses inherent in historical samples across different power stations. Second, the selection of similar days or sequences often employs correlation coefficients, distance metrics, or empirical weights, which can easily introduce historical samples with statistically similar but inconsistent physical driving relationships into the prediction process. Third, under complex weather and extreme fluctuation conditions, effective samples are sparse, making it difficult for single-station models to obtain stable prediction results. Fourth, when deployed across regions, changes in meteorological distribution, capacity scale, and station response relationships can easily lead to significant biases when directly reusing historical samples. Summary of the Invention

[0005] This application provides a method, device, and medium for short-term photovoltaic power prediction based on a causal homogeneous large sample pool, in order to solve the problems of insufficient utilization of single-station sample information, lack of physical constraints for cross-station historical information reuse, susceptibility to statistical correlation misleading in the screening of similar samples, and insufficient prediction stability under complex weather conditions in existing short-term photovoltaic power prediction.

[0006] In a first aspect, this application provides a short-term photovoltaic power prediction method based on a large, homogeneous causal sample pool, the method comprising: Historical measured power, installed capacity, timestamp data, and corresponding numerical weather forecast data of the target photovoltaic power station and multiple reference photovoltaic power stations are obtained. The data of the target photovoltaic power station and the reference photovoltaic power stations are time-aligned, sampling intervals are unified, missing values ​​are handled, and outliers are handled. The measured power is normalized according to the installed capacity of each reference photovoltaic power station and then equivalently converted according to the installed capacity of the target photovoltaic power station to obtain an equivalent power sequence at the same capacity scale as the target photovoltaic power station. Based on the daily power output pattern of the target photovoltaic power station, the time period when the measured power of the target photovoltaic power station is greater than 0MW is taken as the effective power output period. The power change slope of adjacent times is calculated. Based on the power change slope, the stage division rules of the ramp-up period, the steady period and the decay period are determined. The effective power output period of the target photovoltaic power station is marked according to the stage division rules to construct a three-stage target pool. Using the phase division rules of the three-stage target pool as a unified physical boundary, and adopting the same input window length and prediction window length as the target prediction sequence, the complete time series data of the reference photovoltaic power station is sliced ​​by sliding window, and the complete time series is pre-labeled with full timestamp stages to construct a cross-station basic large sample pool. Structural causal models between numerical weather prediction features and photovoltaic power are established for the ramp-up period, the steady period, and the decay period, respectively. The variance of the mean of the potential response of photovoltaic power under different quantile intervention levels for each numerical weather prediction feature is calculated to obtain the intervention importance. The intervention importance is subjected to biased estimation and significance test. The single-site core causal scores of the features that pass the significance test are aggregated across sites to obtain the stable core score. The stable core score is normalized to obtain the normalized comprehensive causal importance score NCCIS. The target photovoltaic power station's stage core feature set is determined based on its NCCIS at the corresponding stage. A cross-station general core feature set is determined based on the NCCIS of multiple reference photovoltaic power stations at the corresponding stage. The general causal matching degree of each reference photovoltaic power station relative to the cross-station general core feature set and the target specific matching degree relative to the target photovoltaic power station's stage core feature set are calculated respectively. The combined weight of the general causal matching degree and the target specific matching degree is determined based on the causal contribution ratio of the general core features to obtain the comprehensive causal matching degree. The stage samples of the reference power stations with a comprehensive causal matching degree lower than a preset threshold are removed to obtain a large sample pool of causal homogeneity. Within the causal homogeneous large sample pool, candidate numerical weather forecast sequences are determined using numerical weather forecast sequences for the target forecast period as the retrieval object. The core feature dimensions of the target sequence and candidate sequences are standardized. The sequence similarity on each core feature dimension is weighted and summed using NCCIS as the weight to obtain multivariate similarity. Based on the multivariate similarity, a set of candidate similar sequences is selected. The illumination amplitude integral is calculated in the set of candidate similar sequences. The historical measured power segment within the time window corresponding to the candidate sequence with the smallest difference from the illumination amplitude integral of the target sequence is selected as the causal homogeneous historical response prior. The target numerical weather forecast sequence, stage labels, NCCIS weights, and the causal homogeneous historical response priors are input into the short-term photovoltaic power prediction model. The core meteorological driving features are enhanced according to the NCCIS weights, the feature interactions of different output stages are constrained according to the stage labels, and the injection intensity of the historical response priors is controlled according to the causal homogeneity to perform gated fusion, thereby obtaining the short-term power prediction results of the target photovoltaic power station.

[0007] In one possible design, the equivalent power sequence of the target photovoltaic power station at the same capacity scale is calculated using the following formula: (1) In the formula: P i ( t ) is the first i individual power station time t The measured power; P rated,i For the first i The installed capacity of each power station For the first i individual power station time t The per-unit value of power; P base The target power plant's installed capacity.

[0008] In one possible design, based on the daily power output pattern of the target photovoltaic power station, the effective power output period is defined as the time period during which the measured power of the target photovoltaic power station is greater than 0 MW. The power change slope between adjacent moments is calculated, and based on the power change slope, the stage division rules for the ramp-up period, the stable period, and the decay period are determined. Specifically, this includes: The effective output period is defined as the time period during which the measured power of the target photovoltaic power station is greater than 0 MW. The slope of the power change between adjacent moments is calculated to characterize the rate of ramp-up or decay of photovoltaic power. The formula for calculating the slope of the power change is as follows: (2) In the formula: k i For the firsti The slope of power change at each moment; P i , P i+1 The first i , i Measured power at time +1; Δt The data sampling time interval; The slope threshold for stage division is determined using the quantile method; the slope threshold for climbing is... k climb and attenuation slope threshold k decay These are defined as the quantiles of all candidate ramp slopes and candidate decay slopes, respectively, and the specific calculation formulas are as follows: (3) In the formula: P peak This represents the maximum power output during the effective output period of a single day; 0.85 P peak The critical power threshold for the end of the ramp-up period and the beginning of the decay period; Based on the ramp slope threshold and the decay slope threshold, a three-stage division rule is determined: power increases from the first time it is greater than 0 MW to the first time it reaches 0.85 MW. P peak The period is the ramp-up phase; power starts from 0.85. P peak The period from the point where the power output drops to the point where it is greater than 0 MW for the last time is the decay period; the remaining period after removing the ramp-up and decay periods from the effective output period is the stable period.

[0009] In one possible design, the phase division rules of the three-stage target pool are used as a unified physical boundary. An input window length and prediction window length consistent with the target sequence to be predicted are used to perform sliding window slicing on the complete time series data of the reference photovoltaic power station. Full timestamp phase pre-labeling is then performed on the complete time series to construct a cross-station basic large sample pool. Specifically, this includes: Using the three-stage partitioning rule determined by the target photovoltaic power station as a unified physical boundary, and employing the same input window length and prediction window length as the target sequence to be predicted, sliding window slicing is performed on the complete time series data of the reference photovoltaic power station, and full timestamp stage pre-labeling is performed on the complete time series to obtain a cross-station basic large sample pool. D 0, the cross-site basic large sample pool D 0 is represented as: (4) In the formula: For the first i Equivalent power sequence of a reference photovoltaic power plant This is a numerical weather prediction sequence; N s The number of photovoltaic power plants is used as a reference.

[0010] In one possible design, structural causal models between numerical weather prediction features and photovoltaic power are established for the ramp-up period, the stable period, and the decay period, respectively. The variance of the mean of the potential response of photovoltaic power for each numerical weather prediction feature at different quantile intervention levels is calculated to obtain the intervention importance. The intervention importance is then subjected to biased estimation and significance testing. The single-site core causal scores of the features that pass the significance test are then aggregated across sites to obtain a stable core score. Finally, the stable core score is normalized to obtain the Normalized Comprehensive Causal Importance Score (NCCIS), specifically including: For the j Numerical weather forecast features X j Apply multiple quantile intervention levels x i (K) The variance of the mean potential response of photovoltaic power under different intervention levels is calculated, and this variance is used as a feature of numerical weather prediction. X j The importance of intervention The calculation formula is as follows: (5) In the formula: Indicates will X j Fixed as the number k Intervention levels; Var k Represents the variance operator with respect to different intervention levels k; Represents the mathematical expectation operator; Y Indicates the photovoltaic power output variable; Dual machine learning was employed to debiased estimate the importance of the interventions. The Bootstrap method was used for significance testing, and the Benjamini-Hochberg false discovery rate was used to correct for and remove non-significant noise features. For features that passed the significance test, the single-station core causality score was calculated using the following formula: (6) In the formula: Features X i In the s The core causal score of a single photovoltaic power station; S s For the first s The set of causal significance features of a photovoltaic power station that passed the significance test; For indicator functions; For the first s Features of a photovoltaic power station X i The estimated importance of the intervention; For the first s Features of a photovoltaic power station X j The estimated importance of the intervention; Stabilized core scores are obtained by merging the individual core causal scores of multiple photovoltaic power plants. : (7) In the formula: For feature X i A collection of sites with non-zero scores. ε To prevent the protection term for logarithmic infinity; if If empty, then StableCore i =0; The stable core score is normalized to obtain the normalized comprehensive causal importance score NCCIS: (8) In the formula: For the first i Numerical weather forecast features X i The normalized comprehensive causal importance score.

[0011] In one possible design, a cross-site common core feature set is determined based on the NCCIS of multiple reference photovoltaic power plants at corresponding stages. The general causal matching degree of each reference photovoltaic power plant relative to the cross-site common core feature set and the target specific matching degree relative to the stage core feature set of the target photovoltaic power plant are calculated using the following formulas: (9) (10) In the formula, For general causal matching degree; Target-specific matching degree; and These represent the target photovoltaic power plant and the reference photovoltaic power plant at different stages. s Lower features X j Normalized comprehensive causal importance score; F core,t A cross-site universal core feature set; F target,t The core feature set for the target photovoltaic power plant stage; This is the core causal feature set of the reference photovoltaic power station j at stage t; The comprehensive causal matching degree is calculated using the following formula: (11) (12) In the formula, For combined weights; To achieve a comprehensive causal fit.

[0012] In one possible design, within the causal homogeneous large sample pool, candidate numerical weather prediction sequences are determined using numerical weather prediction sequences for the target prediction period as the retrieval object. The core feature dimensions of the target sequence and candidate sequences are standardized. A multivariate similarity is obtained by weighted summation of sequence similarities on each core feature dimension using NCCIS as the weight. A set of candidate similar sequences is then selected based on the multivariate similarity. The illumination amplitude integral is calculated within the candidate similar sequence set. The historical measured power segment within the time window corresponding to the candidate sequence with the smallest difference in illumination amplitude integral from the target sequence is selected as the causal homogeneous historical response prior. Specifically, this includes: The formula for calculating multivariate similarity is as follows: (13) In the formula: For multivariate similarity; X , Y These are the target NWP sequence and the candidate NWP sequence, respectively. X i , Y i For the corresponding number i The normalized sequence of class parameters; m is the number of core NWP features involved in similarity calculation; The formula for calculating the integral of the illumination amplitude is as follows: (14) In the formula: I The integral of the illumination amplitude; G ( t ) is the first t Illumination amplitude value; Δt For time step; n This represents the total time step for the corresponding stage.

[0013] In one possible design, gated fusion is performed based on the injection intensity of the historical response prior, controlled by causal homogeneity, to obtain the short-term power prediction results of the target photovoltaic power station, specifically including: Gated fusion is performed based on the injection strength of the historical response prior, controlled by causal homogeneity, with the gating fusion coefficient...g The calculation formula is as follows: (15) Where: NCCIS pool Mean pooling for all core feature NCCIS scores; W is the Sigmoid activation function. g is the learnable weight matrix of the gated fusion layer; CosSim is the NCCIS weighted cosine similarity between the target NWP sequence and the matching historical NWP sequence, used to characterize the degree of causal homogeneity between the two; b g For learnable bias terms of the gated fusion layer; Based on the gating fusion coefficient, the short-term power prediction result of the target photovoltaic power station is calculated using the following formula: (16) In the formula: P NWP These are the base predictions obtained by the prediction model based on the target NWP sequence; For the short-term predicted power of the target photovoltaic power plant; The causal homogeneity between the target scene and the matched historical scene is represented by the NCCIS weighted cosine similarity. When the causal homogeneity between the target scene and the historical scene is higher than a preset threshold, the gating coefficient is adjusted. g As the causal homogeneity increases, the model relies more heavily on prior causal homogeneity of historical responses to correct for NWP system bias; when the causal homogeneity is below a preset threshold, the gating coefficient... g The model is reduced, and it relies more on the underlying predictions obtained by NWP.

[0014] Secondly, this application provides a short-term photovoltaic power prediction device based on a causal homogeneous large sample pool, the device comprising: The data preprocessing module is configured to acquire historical measured power, installed capacity, timestamp data, and corresponding numerical weather forecast data of the target photovoltaic power station and multiple reference photovoltaic power stations. It performs time alignment, sampling interval unification, missing value processing, and outlier processing on the data of the target photovoltaic power station and the reference photovoltaic power stations. After capacity normalization of the measured power based on the installed capacity of each reference photovoltaic power station, it performs equivalent conversion according to the installed capacity of the target photovoltaic power station to obtain an equivalent power sequence at the same capacity scale as the target photovoltaic power station. The target pool construction module is configured to calculate the power change slope between adjacent times based on the daily power output pattern of the target photovoltaic power station, taking the time period when the measured power of the target photovoltaic power station is greater than 0MW as the effective power output period, and determine the stage division rules of the ramp-up period, the steady period and the decay period based on the power change slope. The effective power output period of the target photovoltaic power station is marked according to the stage division rules to construct a three-stage target pool. The basic sample construction module is configured to use the stage division rules of the three-stage target pool as the unified physical boundary, adopt the input window length and prediction window length consistent with the target prediction sequence, perform sliding window slicing on the complete time series data of the reference photovoltaic power station, and perform full timestamp stage pre-labeling on the complete time series to construct a cross-station basic large sample pool. The scoring calculation module is configured to establish structural causal models between numerical weather forecast features and photovoltaic power for the climbing period, the stable period, and the decay period, respectively. It calculates the variance of the mean of the potential response of photovoltaic power under different quantile intervention levels for each numerical weather forecast feature to obtain the intervention importance. It performs biased estimation and significance test on the intervention importance. It performs cross-site stabilization aggregation on the single-site core causal scores of the features that pass the significance test to obtain the stable core score. It performs normalization processing on the stable core score to obtain the normalized comprehensive causal importance score NCCIS. The causal homogeneous sample construction module is configured to determine the stage core feature set of the target photovoltaic power station based on the NCCIS of the target photovoltaic power station at the corresponding stage, determine the cross-station general core feature set based on the NCCIS of multiple reference photovoltaic power stations at the corresponding stage, calculate the general causal matching degree of each reference photovoltaic power station relative to the cross-station general core feature set and the target specific matching degree relative to the stage core feature set of the target photovoltaic power station, determine the combined weight of the general causal matching degree and the target specific matching degree based on the causal contribution ratio of the general core features to obtain the comprehensive causal matching degree, and remove the stage samples of the reference power station with the comprehensive causal matching degree lower than the preset threshold to obtain a large causal homogeneous sample pool. The response prior calculation module is configured to determine candidate numerical weather forecast sequences within the causal homogeneous large sample pool, using numerical weather forecast sequences of the target forecast period as the retrieval object; standardize the core feature dimensions of the target sequence and candidate sequences; use NCCIS as the weight to perform a weighted summation of the sequence similarity on each core feature dimension to obtain a multivariate similarity; filter the candidate similar sequence set based on the multivariate similarity; calculate the illumination amplitude integral in the candidate similar sequence set; and select the historical measured power segment within the time window corresponding to the candidate sequence with the smallest difference from the illumination amplitude integral of the target sequence as the causal homogeneous historical response prior. The short-term power prediction module is configured to input the target numerical weather forecast sequence, stage labels, NCCIS weights, and the causal homogeneous historical response priors into the short-term photovoltaic power prediction model. It enhances the core meteorological driving features according to the NCCIS weights, constrains the feature interactions of different output stages according to the stage labels, and performs gated fusion by controlling the injection intensity of the historical response priors according to the causal homogeneity to obtain the short-term power prediction results of the target photovoltaic power station.

[0015] Thirdly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions. When a processor executes the computer-executable instructions, it implements the short-term photovoltaic power prediction method based on a causal homogeneous large sample pool as described in the first aspect and various possible designs of the first aspect.

[0016] The short-term photovoltaic power prediction method, device, and dielectric provided in this application based on a causal homogeneous large sample pool have at least the following beneficial effects: This application starts with the intraday power output patterns of photovoltaic power plants and the need for cross-site historical sample reuse, and proposes a method that can constrain the scope of historical sample reuse using causal structure. Through a three-stage target pool construction, NCCIS causal importance scoring, purification of a large causal homogeneous sample pool, similar NWP sequence matching, and controllable injection of historical response priors, similar historical power is no longer a simple splicing of additional features, but a reusable historical response prior after causal boundary constraints. This improves the stability, interpretability, and cross-regional adaptability of short-term photovoltaic power prediction. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] Figure 1 Provided for the embodiments of this application; Figure 2 A flowchart illustrating a short-term photovoltaic power prediction method based on a causal homogeneous large sample pool, provided for embodiments of this application; Figure 3 A schematic diagram of the power output distribution and three-stage physical boundaries of the target power plant in different seasons, provided for an embodiment of this application. Figure 4 A schematic diagram of the structure of the historical response prior and NCCIS constraint injection prediction model provided for embodiments of this application; Figure 5 This is a structural diagram of a short-term photovoltaic power prediction device based on a causal homogeneous large sample pool, provided in an embodiment of this application.

[0019] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concepts of this application to those skilled in the art through reference to specific embodiments. Detailed Implementation

[0020] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0021] The collection, storage, use, processing, transmission, provision, and disclosure of relevant data and information in the technical solution of this application all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0022] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.

[0023] The technical solution of this application and how it solves the above-mentioned technical problems will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will be described below with reference to the accompanying drawings.

[0024] Example 1: This application provides a short-term photovoltaic power prediction method based on a causal homogeneous large sample pool. This method is used for short-term photovoltaic power prediction by constructing a causal homogeneous large sample pool, filtering historical power segments consistent with the physical drivers of the target scenario, and using these as historical response priors in the prediction, thereby improving the prediction stability and transferability under complex weather and cross-site reuse scenarios. Figure 1 As shown in the embodiments of this application, this is the overall technical roadmap for short-term photovoltaic power prediction based on a large sample pool of causal homogeneity. The implementation of this method is based on a data layer of multi-source data and physical stage boundaries, a mechanism layer for NCCIS causal structure identification, a core mechanism layer for reusing causal homogeneous historical information, and a prediction model layer.

[0025] The multi-source data and physical stage boundary data layer serves as the basic input unit, comprising four types of input elements. The first type is the target station NWP data, covering meteorological elements such as shortwave radiation, cloud cover, and temperature, providing meteorological driving input for power prediction. The second type is the target station's historical power data, consisting of a 15-minute resolution output sequence and corresponding installed capacity data, used to construct a benchmark sample and output stage division for the target power station. The third type is a cross-power station historical sample set, containing historical measured power and corresponding NWP data from multiple reference photovoltaic power stations, serving as the core data source for constructing a large cross-station sample pool and expanding available historical information. The fourth type is a three-stage label, corresponding to the output stage division results for the ramp-up period, stable period, and decay period, providing a unified physical boundary for subsequent hierarchical causal identification and staged prediction.

[0026] The role of NCCIS causal structure identification is to determine the "reusable boundary" and "reusable weight" of cross-site samples, belonging to the mechanism layer of the method. Based on multi-site power and NWP data input from the data layer, this NCCIS causal structure identification performs causal feature identification for the ramp-up, plateau, and decay periods respectively: First, a structural causal model is constructed and the intervention importance is calculated. By applying quantile intervention to various numerical weather forecast features, the causal driving strength of different meteorological elements on photovoltaic power is quantified. Then, dual machine learning is used for debiased estimation to eliminate the interference of confounding variables on the causal effect estimation. Next, through Bootstrap significance testing and false discovery rate correction, non-significant noise features are eliminated, retaining the core driving elements with stable causal effects. Based on this, the single-site core causal scores of multiple power plants are stabilized and aggregated using StableCore to filter single-site statistical noise and extract the cross-site common meteorological-power physical driving laws. Finally, the normalized comprehensive causal importance score NCCIS is calculated. This score will be used for subsequent causal homogeneity purification of the sample pool, weight calculation for similar sequence matching, and prior injection strength control in the prediction fusion stage.

[0027] The causal homogeneous historical information reuse mechanism belongs to the model layer and is also the core mechanism layer of the method. Based on the upper-layer NCCIS score and cross-site samples in the data layer, it completes the construction of a large causal homogeneous sample pool and the extraction of historical response priors in three steps. The first step is causal homogeneous purification. According to the causal feature set of each reference power station at the corresponding stage and the NCCIS score, the general causal matching degree and the target specific matching degree are calculated respectively, and the comprehensive causal matching degree is obtained by fusion. Cross-site samples with similar statistics but inconsistent physical driving relationships are eliminated, and only reference power station samples with causal homogeneity with the target scenario are retained, thus completing the purification of the cross-site basic large sample pool. The second step is causal weight matching. In the causal homogeneous large sample pool, the NWP sequence of the target prediction period is used as the retrieval object, and the multivariate sequence similarity is calculated with NCCIS as the feature weight. The initial screening obtains a subset of similar sequences with similar meteorological conditions; at the same time, the irradiance amplitude integral is introduced for secondary screening to ensure the matching consistency of the total irradiance energy dimension and further narrow the candidate range. The third step is historical response prior extraction. From the time window corresponding to the most similar NWP sequence obtained through screening, the corresponding historical measured power segment is extracted to obtain the causal homogeneous historical response prior P. sim This prior is not a simple concatenation of similar power outputs, but rather a reusable historical output response constrained by causal boundaries, which can be used to correct systemic biases in NWP-driven predictions. This module synchronously outputs the matched NWP sequence and NCCIS weights, which are then fed into subsequent prediction modules.

[0028] The prediction model layer is adaptable to various short-term power prediction model architectures such as DDT-PV, and receives multiple types of information output from the core mechanism layer to complete the final power prediction. This prediction model layer includes five core processing operations: One is the three-stage decoupling expert, which constrains the characteristic interaction patterns of different output stages based on the physical boundaries of the ramp-up period, the steady period, and the decay period, so that the prediction process conforms to the physical laws of photovoltaic intraday output. The second is causal attention coding, which encodes the temporal NWP features and power features to capture the temporal and variable correlations between meteorological drives and power responses; Thirdly, NCCIS prior injection embeds the causal importance score into the attention layer and output layer, strengthening the contribution ratio of core meteorological driving features and suppressing the interference of non-core noise features. Fourthly, gated causal fusion calculates the gating coefficient g based on the causal homogeneity between the target scene and historical scenes, and controllably gating the causal homogeneous historical response prior P. sim The model is weighted and fused with the basic prediction value PNWP obtained from NWP. When the causal homogeneity is high, the prior weight is increased to correct the systematic bias of NWP. When the homogeneity is low, the model relies more on the basic prediction results driven by NWP. Fifthly, Bayesian online correction is used to perform secondary calibration on the prediction bias from both the time and frequency domains, further improving the stability and accuracy of the prediction results. Finally, the model outputs the short-term power prediction results for the target photovoltaic power plant, completing the entire prediction process.

[0029] Based on the above technical approach, please further combine... Figure 2 As shown, the short-term photovoltaic power prediction method based on a causal homogeneous large sample pool is executed according to the following steps S10-S70.

[0030] S10: Obtain historical measured power, installed capacity, timestamp data, and corresponding numerical weather forecast data for the target photovoltaic power station and multiple reference photovoltaic power stations. Perform time alignment, sampling interval unification, missing value processing, and outlier processing on the data of the target photovoltaic power station and reference photovoltaic power stations. After capacity normalization of the measured power based on the installed capacity of each reference photovoltaic power station, perform equivalent conversion according to the installed capacity of the target photovoltaic power station to obtain an equivalent power sequence at the same capacity scale as the target photovoltaic power station.

[0031] Step S10 aims to perform basic data preprocessing and capacity equivalence conversion. Specifically, it acquires historical measured power, installed capacity, timestamp data, and corresponding numerical weather forecast data for the target photovoltaic power station and multiple reference photovoltaic power stations. The numerical weather forecast data includes one or more of shortwave radiation, temperature, humidity, wind speed, cloud cover, and air pressure. The data from the target and reference photovoltaic power stations are then time-aligned, sample intervals are standardized, missing values ​​are handled, and outliers are processed.

[0032] Because different photovoltaic power plants have different installed capacities, directly using the original power sequences to construct a cross-site sample pool can lead to capacity scale differences affecting the matching of similar sequences. To eliminate the impact of installed capacity differences on cross-site sample reuse, the measured power of each reference photovoltaic power plant is normalized by capacity and then equivalently converted according to the installed capacity of the target photovoltaic power plant to obtain an equivalent power sequence at the same capacity scale as the target photovoltaic power plant. The specific formula is as follows: (1) In the formula: P i ( t ) is the first i individual power station time t The measured power; P rated,i For the first i The installed capacity of each power station For the first i individual power station time t The per-unit value of power; P base The target power plant's installed capacity.

[0033] S20: Based on the daily power output pattern of the target photovoltaic power station, the effective power output period is defined as the time period when the measured power of the target photovoltaic power station is greater than 0MW. The power change slope between adjacent times is calculated. Based on the power change slope, the stage division rules for the ramp-up period, the steady period and the decay period are determined. The effective power output period of the target photovoltaic power station is marked according to the stage division rules to construct a three-stage target pool.

[0034] Photovoltaic power output exhibits a distinct diurnal cycle, roughly encompassing a ramp-up output period, a stable output period, and a decline output period under sunny, cloudy, and fluctuating weather conditions. To enable the reuse of historical samples from different photovoltaic power stations within the same physical boundaries, this embodiment constructs a three-stage target pool based on the intraday output pattern of the target photovoltaic power station through step S20.

[0035] In some embodiments, the time period during which the measured power of the target photovoltaic power station is greater than 0 MW is taken as the effective output period, and the slope of the power change between adjacent time points is calculated to characterize the rate of ramp-up or decay of photovoltaic power. The specific formula is as follows: (2) In the formula: k i For the first i The slope of power change at each moment, in MW / min; P i , P i+1 The first i , i The measured power at time +1 is expressed in MW. Δt =15min is the data sampling time interval.

[0036] The quantile method is used to determine the slope threshold for stage division. (Climbing slope threshold) k climb and attenuation slope threshold k decay These are defined as the quantiles of all candidate ramp slopes and candidate decay slopes, respectively, as shown in the following formulas: (3) In the formula: P peak This represents the maximum power output during the effective output period of a single day, in MW; 0.85 P peak This is the critical power threshold at the end of the ramp-up period and the beginning of the decay period.

[0037] Based on the above thresholds, a three-stage division rule is determined: power increases from the first time it exceeds 0 MW to the first time it reaches 0.85 MW. P peakThe period is the ramp-up phase; power starts from 0.85. P peak The period from when the output drops to the last time it exceeds 0 MW is the decay period; the remaining period after excluding the ramp-up and decay periods from the effective output period is the stable period. Based on the stage labels of the ramp-up, stable, and decay periods, a three-stage target pool for the target photovoltaic power plant is constructed.

[0038] like Figure 3 The figure shows a schematic diagram of the daily power output distribution of the target photovoltaic power station in different seasons according to an embodiment of this application. The figure displays the photovoltaic power output curves for multiple natural days in spring, summer, autumn, and winter. The vertical axis represents photovoltaic power generation in MW; the horizontal axis represents the daily time step, with each time step corresponding to a 15-minute data sampling interval. Each curve corresponds to the actual power output sequence of the target photovoltaic power station for one natural day. By superimposing multiple daily power output curves, the daily variation characteristics of photovoltaic power under different seasons and weather conditions can be reflected.

[0039] Depend on Figure 3 It is evident that the daily photovoltaic (PV) output varies across different seasons in terms of specific power levels, effective output duration, and fluctuations. However, it generally exhibits a phased evolutionary pattern: a gradual increase from low output, entry into the mid-day operating phase, and subsequent gradual decline to near zero output. This pattern aligns with the intraday variations in solar irradiance and solar altitude angle, providing a data foundation and physical basis for dividing the daily effective PV output process into three physical stages: a ramp-up period, a stable period, and a decline period.

[0040] The effective power output period varies significantly across different seasons. Summer has a relatively long effective power output period, while winter has a relatively short one, with spring and autumn falling in between. This reflects the impact of seasonal variations on the available irradiance duration and intraday power output of photovoltaic systems. Furthermore, there are significant differences in power output curves between different natural days. Under relatively stable meteorological conditions, the intraday power output curve is generally smooth and approaches a high-output envelope; however, under the influence of changes in cloud cover, irradiance fluctuations, and other meteorological conditions, the power output curve may exhibit varying degrees of local fluctuations and power reductions. Therefore, although photovoltaic power exhibits a stable overall intraday evolution pattern, significant random fluctuations still exist within each stage.

[0041] Based on the aforementioned intraday power output characteristics, this application employs a method combining power change slope, intraday peak power, and quantile statistics to determine the three-stage physical boundaries. Specifically, the effective power output process begins when the power first exceeds 0 MW, and the maximum power output during the single-day effective power output period is used as the starting point. P peak Based on the benchmark: the ramp-up period begins when the power output first exceeds 0 MW, and the overall output gradually increases from low to high until it first reaches 0.85 MW.P peak The decay period is characterized by the overall decline in power output from high to low levels, starting from 0.85. P peak The power output decreases until it reaches a maximum of 0 MW; the stable period is the intermediate operating phase after excluding the ramp-up and decay periods within the effective output period. Due to weather changes, a certain degree of local power fluctuation is allowed within each phase, and a strictly monotonic power change is not required.

[0042] Furthermore, addressing the differences in sunrise and sunset times and effective power output duration across different seasons, this application uses the statistical distribution of the start and end times of the three stages for each natural day, employing the quantile method to determine the typical time range of each stage under different seasons, and calibrates it to a 15-minute time resolution consistent with the original data. This results in a three-stage physical boundary that adaptively adjusts with seasonal changes, rather than using a fixed time to uniformly and rigidly divide the annual photovoltaic power output, thus better adapting to the differences in intraday power output patterns under different seasons and weather conditions.

[0043] By dividing the process into three stages, a unified physical stage label can be assigned to the target power station and cross-station historical samples. This allows historical samples from different power stations, different seasons, and different days to be aligned according to the same photovoltaic power output evolution stage. This provides a unified physical boundary and sample organization basis for subsequent staged NCCIS causal structure identification, causal homogeneity purification of cross-station sample pools, similar NWP sequence matching, and prior extraction of causal homogeneity historical responses. It also reduces the risk of mismatch of historical samples caused by seasonal differences, power output duration differences, and stage misalignment.

[0044] S30: Using the phase division rules of the three-stage target pool as a unified physical boundary, and adopting the same input window length and prediction window length as the target prediction sequence, the complete time series data of the reference photovoltaic power station is sliced ​​by sliding window, and the complete time series is pre-labeled with full timestamp stages to construct a cross-station basic large sample pool.

[0045] To expand the available historical samples while preserving the continuous time-series structure of photovoltaic power and numerical weather prediction data, this embodiment constructs a cross-station basic large sample pool in step S30. This cross-station basic large sample pool does not simply segment the original time series, but rather performs full-timestamp pre-annotation on the complete time series, and then generates samples using a sliding window approach.

[0046] In some embodiments, the specific method for constructing a cross-site basic large sample pool is as follows: using the three-stage division rule determined by the target photovoltaic power station as a unified physical boundary, and adopting the same input window length and prediction window length as the target sequence to be predicted, the complete time series data of the reference photovoltaic power station is sliced ​​by sliding window.

[0047] Cross-site basic large sample pool D 0 can be represented as: (4) In the formula: For the first i Equivalent power sequence of a reference photovoltaic power plant Numerical weather forecast sequence ; N s The number of photovoltaic power plants is used as a reference.

[0048] S40: Establish structural causal models between numerical weather prediction features and photovoltaic power for the ramp-up, steady-state, and decay periods respectively. Calculate the variance of the mean of the potential response of photovoltaic power for each numerical weather prediction feature at different quantile intervention levels to obtain intervention importance. Perform biased estimation and significance testing on the intervention importance. Stabilize and aggregate the single-site core causal scores of features that pass the significance test across sites to obtain stable core scores. Normalize the stable core scores to obtain the normalized comprehensive causal importance score (NCCIS).

[0049] To overcome the problem that traditional correlation feature screening is easily affected by spurious correlations, this embodiment establishes structural causal models between numerical weather forecast features and photovoltaic power for the ramp-up period, the steady period, and the decline period, respectively, and calculates the normalized comprehensive causal importance score (NCCIS) to identify the core meteorological driving factors that stably affect photovoltaic output at different stages.

[0050] In some embodiments, for the first j Numerical weather forecast features X j Apply multiple quantile intervention levels x i (K) The variance of the mean potential response of photovoltaic power under different intervention levels was calculated, and this variance was used as a feature. X j The importance of intervention The specific formula is as follows: (5) In the formula: Indicates will X i Fixed as the number k Intervention levels; Var k Represents the variance operator with respect to different intervention levels k; Represents the mathematical expectation operator; Y This represents the photovoltaic power output variable. Π( X i The larger the value, the greater the change.X i The more severe the power fluctuations caused, the stronger the causal effect of this feature.

[0051] A dual machine learning approach was employed to debiased estimate the importance of interventions. The Bootstrap method was used for significance testing, and the Benjamini-Hochberg false discovery rate was used to correct for and remove non-significant noise features. For features that passed the significance test, their single-station core causality score was calculated using the following formula: (6) In the formula: Features X j In the s The core causal score of a single photovoltaic power station; S s For the first s The set of causal significance features of a photovoltaic power station that passed the significance test; This is an indicator function.

[0052] To filter out statistical noise from individual photovoltaic power plants and preserve stable meteorological-power physical driving forces across plants, the core causal scores of individual plants from multiple photovoltaic power plants are stabilized and aggregated to obtain stable core scores. The specific formula is as follows: (7) In the formula: For feature X i A collection of sites with non-zero scores. ε =10 -8 To prevent the protection term for logarithmic infinity. If If empty, then StableCore i =0.

[0053] Finally, the stable core scores are normalized to obtain the normalized comprehensive causal importance score NCCISi, as shown in the following formula: (8) NCCIS is used for subsequent sample pool purification, similar sequence matching, and historical response prior injection weight control.

[0054] S50: Determine the stage core feature set of the target photovoltaic power station based on the NCCIS of the target photovoltaic power station at the corresponding stage. Determine the cross-station general core feature set based on the NCCIS of multiple reference photovoltaic power stations at the corresponding stage. Calculate the general causal matching degree of each reference photovoltaic power station relative to the cross-station general core feature set and the target specific matching degree relative to the stage core feature set of the target photovoltaic power station. Determine the combined weight of the general causal matching degree and the target specific matching degree based on the causal contribution ratio of the general core features to obtain the comprehensive causal matching degree. Remove the stage samples of the reference power station with a comprehensive causal matching degree lower than the preset threshold to obtain a large sample pool of causal homogeneity.

[0055] This embodiment uses NCCIS to perform causal homogeneity purification on a large cross-site basic sample pool, screening reference power stations and historical samples that have consistent physical driving relationships with the target photovoltaic power station at the corresponding stage, in order to construct a large causal homogeneity sample pool. The purpose of causal homogeneity purification is not simply to screen samples with the closest statistical distance, but to exclude historical samples that are statistically similar but have inconsistent physical driving relationships.

[0056] In some embodiments, the core feature set of the target photovoltaic power plant stage is determined based on the NCCIS of the target photovoltaic power plant at the corresponding stage. F target,t Determine the cross-site common core feature set based on the NCCIS of multiple reference photovoltaic power plants at the corresponding stages. F core,t The general causal matching degree of each reference photovoltaic power station relative to the cross-station general core feature set, and the target specific matching degree relative to the target photovoltaic power station's stage core feature set are calculated separately, with the specific formulas as follows: (9) (10) In the formula ， General causal matching degree ； Target-specific matching degree ； and These represent the target photovoltaic power plant and the reference photovoltaic power plant at different stages. s Lower features X j The normalized comprehensive causal importance score.

[0057] The combined weights of the two are determined based on the causal contribution ratios of the common core features, thus obtaining the comprehensive causal matching degree. The specific formula is as follows: (11) (12) In the formula, For combined weights.

[0058] S60: Within a large, homogeneous causal sample pool, candidate numerical weather prediction sequences are determined using numerical weather prediction sequences for the target prediction period as the retrieval object. The core feature dimensions of the target sequence and candidate sequences are standardized. The sequence similarity on each core feature dimension is weighted and summed using NCCIS as the weight to obtain multivariate similarity. Based on multivariate similarity, a set of candidate similar sequences is selected. The illumination amplitude integral is calculated in the set of candidate similar sequences. The historical measured power segment within the time window corresponding to the candidate sequence with the smallest difference from the illumination amplitude integral of the target sequence is selected as the causal homogeneous historical response prior.

[0059] Similar numerical weather forecast sequences are matched within a large, homogeneous causal sample pool. Using numerical weather forecast sequences for the target forecast period as the retrieval object, candidate numerical weather forecast sequences are identified within the large, homogeneous causal sample pool, and the core feature dimensions of the target and candidate sequences are standardized.

[0060] In some embodiments, the sequence similarity between the target numerical weather prediction sequence and the candidate numerical weather prediction sequence is calculated on each core feature dimension, and a weighted sum is performed using NCCIS as the weight to obtain the multivariate similarity, as shown in the following formula: (13) In the formula: X , Y These are the target NWP sequence and the candidate NWP sequence, respectively. X i , Y i For the corresponding number i Normalized sequence of class parameters; MultiVarSim( X , Y )∈[0,1], the larger the value, the more similar the meteorological conditions corresponding to the two sequences are, forming a subset of similar sequences.

[0061] In the subset, the illumination amplitude sequence within the target NWP sequence and candidate NWP sequences is used. Based on the time step and length of the short-term predicted sequence, the integral of the illumination amplitude within the time period is calculated to characterize the total irradiance received per unit area. The calculation formula is as follows: (14) In the formula: I Integral of illumination amplitude (unit: kW·h / m) 2 ); G(t) For the first t Step illumination amplitude value; Δt For time step; nThis represents the total time step for the corresponding stage.

[0062] Calculate the integral of illumination amplitude for each target sequence. I target And the integral of illumination amplitude for each sequence within the matched candidate subset. I can,j The historical measured power segments within the time window corresponding to the most similar sequences were selected and identified as causal homogeneous historical response priors. P sim .

[0063] S70: Input the target numerical weather forecast sequence, stage labels, NCCIS weights, and causal homogeneous historical response priors into the short-term photovoltaic power prediction model. Enhance the core meteorological driving features according to the NCCIS weights, constrain the feature interactions of different output stages according to the stage labels, and perform gating fusion according to the injection intensity of historical response priors controlled by causal homogeneity to obtain the short-term power prediction results of the target photovoltaic power plant.

[0064] In some embodiments, the causal homogeneity between the target scene and the matching historical scene is characterized by the NCCIS weighted cosine similarity. Specifically, based on the Normalized Comprehensive Causal Importance Score (NCCIS) corresponding to each core NWP feature, the standardized sequences of corresponding NWP features in the target scene and the matching historical scene are weighted, and the causal homogeneity is calculated. In the formula, The degree of causal homogeneity between the target scene and the matched historical scenes; m The number of core NWP features involved in calculating causal homogeneity; For the first i Normalized comprehensive causal importance score of the core NWP features; and These are the target scene and the matching historical scene, respectively. i The standardized feature values ​​or feature sequences of each NWP feature. The greater the degree of causal homogeneity, the more consistent the target scenario and the matching historical scenario are in terms of meteorological driving features that have a significant causal contribution to photovoltaic power, and the higher the reference value of the power response of the matching historical scenario for the target scenario; conversely, it indicates that the main meteorological driving mechanisms of the two are less consistent, and the reference value of the corresponding historical power response is lower.

[0065] In some embodiments, the target numerical weather prediction sequence, stage labels, NCCIS weights, and causally homogeneous historical response priors are input into a short-term photovoltaic power prediction model. In the prediction model, core meteorological driving features are enhanced based on the NCCIS weights, feature interactions at different power output stages are constrained based on the stage labels, and the injection intensity of historical response priors is controlled based on causal homogeneity. Gated fusion coefficients are used. g The calculation formula is as follows: (15) Where: NCCIS pool Mean pooling for all core feature NCCIS scores; W g is the learnable weight matrix of the gated fusion layer; CosSim is the NCCIS weighted cosine similarity between the target NWP sequence and the matching historical NWP sequence, used to characterize the degree of causal homogeneity between the two; b g This is a learnable bias term for the gated fusion layer.

[0066] The final predicted value is shown in the following formula: (16) In the formula: P NWP This is the base prediction value obtained by the prediction model based on the target NWP sequence; the prediction model can employ a Transformer encoder, a temporal neural network, an ensemble learning model, or other regression models capable of short-term power prediction. The comprehensive causal matching degree is used as the causal homogeneity; when the causal homogeneity between the target scene and historical scenes is high, the gating coefficient... g As the causal homogeneity increases, the model relies more heavily on causal homogeneity of historical response priors to correct for NWP system bias; when causal homogeneity is low, the gating coefficient... g The model relies more on the underlying predictions obtained by NWP.

[0067] like Figure 4 The diagram shown is a structural schematic of the historical response prior and NCCIS constraint-injected prediction model provided in this embodiment of the application. The prediction model adopts a hierarchical modular architecture design, consisting of an input layer, an encoding layer, and an output layer from top to bottom. Finally, after Bayesian online correction, it outputs short-term photovoltaic power prediction results, achieving controllable fusion of meteorological driving feature encoding under causal constraints and historical response prior.

[0068] The input layer provides three core inputs to the model: NCCIS weighted NWP sequences, three-stage labels, and historical response priors. The NCCIS weighted NWP sequences are time-series sequences obtained by weighting numerical weather prediction features according to the normalized comprehensive causal importance score of the corresponding stage, used to strengthen the contribution ratio of core meteorological driving features at the input end. The three-stage labels correspond to the stage division results of the climbing period, the stable period, and the decay period, providing a unified physical boundary identifier for the staged decoupling processing of the coding layer. The historical response priors are P values ​​obtained through a large pool of homogeneous causal samples. sim As a reusable historical output reference verified by physical causality, it is input to the gating fusion unit of the output layer.

[0069] The encoding layer is a three-stage decoupled dual-expert causal constraint Transformer encoding layer, which is the core unit for the model to extract temporal features and inject causal constraints. This encoding layer has independent encoders for the three physical stages of photovoltaic power output: a ramp-up encoder, a steady-state encoder, and a decay encoder. These three types of encoders process the input features of their respective stages in parallel, avoiding interference between the physical laws of different output stages. Each encoder internally includes the following four types of functional units: 1) SCM Causal Attention Mask: Based on the causal dependencies of the structural causal model, it imposes constraints on attention calculation, restricts the attention weights of non-causal features, and avoids interference from spurious correlations at the feature encoding level. 2) NCCIS prior injection embeds the normalized comprehensive causal importance score into the attention weight calculation process, further increasing the attention ratio of core driving features; 3) Dual expert mechanism, setting up parallel feature extraction paths to adapt to different weather complexities, taking into account both stable power output on sunny days and complex power output characteristics on cloudy and fluctuating weather. 4) Local control fusion is used to adaptively balance the output characteristics of each unit within the encoder and achieve optimal feature fusion within a stage.

[0070] The three types of encoders output the encoded features of their respective stages. h climb , h stable , h decay They are all input to the global fusion module of the output layer.

[0071] The output layer consists of three parts: a global fusion module H, a prediction projection layer, and a gated causal fusion module. The global fusion module H receives the encoded features from the three stages, performs cross-stage global feature integration and temporal correlation modeling, and forms a global feature representation covering the complete power output process. The prediction projection layer maps the global fusion features to power prediction values, namely the basic prediction value PNWP obtained based on the target NWP sequence, which corresponds to the baseline prediction result derived only from numerical weather prediction features.

[0072] The gated causal fusion module is the core unit for realizing the controllable prior injection of historical responses, based on the basic predicted values. P NWP Homogeneous historical response priors with causality P sim As two inputs, the gated causal fusion module calculates the causal homogeneity between the target predicted scene and the historical prior scene, and generates gating coefficients through an MLP network. g The calculation of the gating coefficients simultaneously incorporates the mean pooling result of causal homogeneity and the core feature NCCIS score; adaptive fusion of the two signals is performed according to the weighted fusion formula to obtain the fused prediction value. P pred When the causal homogeneity between the target scene and the historical scene is high, the gating coefficient... g As the model increases, it relies more on historical response priors that have been causally verified to correct the systematic bias of NWP; when the causal homogeneity is low, the gating coefficient... g To reduce the bias, the model is based primarily on NWP-driven prediction results, avoiding the introduction of prediction bias by heterogeneous historical information and ensuring the stability and rationality of the prediction results.

[0073] The preliminary prediction results obtained through gated causal fusion are further input into the Bayesian online correction module for real-time error calibration. Combined with the latest measured power output data, the prediction bias is dynamically adjusted to improve prediction accuracy under complex fluctuating weather and meteorological abrupt changes. The output of the prediction model is the short-term power prediction result P of the target photovoltaic power station. Its time-series resolution is consistent with the input data, providing direct data support for power system dispatching and operation, reserve capacity configuration, and renewable energy consumption decisions. The overall model architecture uses causal consistency as the core constraint. It adapts to the physical laws of different power output stages through phased decoupling coding and achieves controllable injection of historical priors through a gating mechanism. This fully exploits the effective information of homogeneous causal samples across stations while effectively avoiding prediction errors caused by heterogeneous samples.

[0074] In some embodiments, after step S70, the short-term photovoltaic power prediction method based on a causal homogeneous large sample pool further includes the following steps S80 and S90.

[0075] S80: Simulation calculation.

[0076] Simulation inputs include: historical measured power data of the target photovoltaic power station of sufficient length; historical measured power data of multiple reference photovoltaic power stations; numerical weather forecast data of the target and reference photovoltaic power stations at corresponding times; installed capacity of the target and reference photovoltaic power stations; and a data sampling interval of 15 minutes. Based on steps 1) to 7), the short-term power prediction results of the target photovoltaic power station are obtained.

[0077] S90: Error Analysis.

[0078] Select the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R-squared, R0). 2 Prediction accuracy (ACC) and prediction accuracy are used as performance evaluation metrics for prediction models, and their calculation formulas are as follows: (17) (18) (19) (20) In the formula: N is the total number of samples in the test set. This is the measured value of photovoltaic power. These are the model's predicted values. The installed capacity of the target photovoltaic power station.

[0079] Example 2: Based on the short-term photovoltaic power prediction method based on a causal homogeneous large sample pool provided in Example 1, this example provides a detailed explanation of the feasibility and advancement of the method by combining specific experimental data.

[0080] In this embodiment, the short-term photovoltaic power prediction method based on a causal homogeneous large sample pool is implemented through the following steps 1-9. It should be noted that the specific calculation process involved is consistent with the calculation method given in Embodiment 1.

[0081] Specifically, the method of this application is applicable to short-term power prediction of photovoltaic power plants. Multiple reference photovoltaic power plants can be used; in this embodiment, 19 photovoltaic power plants in a certain region are used as the source of the cross-site sample pool, and one of the target photovoltaic power plants is used as the prediction object for experimental verification. The data can be divided into training and test sets in chronological order; in this embodiment, the training set accounts for 80%, and the test set accounts for 20%. Photovoltaic power plant sample information is shown in Table 1.

[0082] Table 1. Sample Information of Photovoltaic Power Plants

[0083] Step 1: Read the historical measured power, installed capacity, timestamp, and corresponding NWP data of the target photovoltaic power station and multiple reference photovoltaic power stations; in this embodiment, there are 19 reference photovoltaic power stations. The data of each power station are time-aligned and missing values ​​are processed according to a uniform 15-minute time resolution, and the power of the reference photovoltaic power stations is converted to the capacity scale of the target photovoltaic power station using formulas (1) and (2).

[0084] Step 2: Using the historical power sequence of the target photovoltaic power station as the object, calculate the slope of power change at adjacent times according to formula (3), and then determine the physical boundaries of the ramp-up period, the steady period, and the decay period according to formulas (4) and (5). Label the target sequence to be predicted according to its stage to obtain the three-stage target pool.

[0085] Step 3: Using the three-stage division results of the target photovoltaic power station as a unified physical benchmark, pre-label the complete historical time series data of the reference photovoltaic power station with full timestamps. Then, a large sample pool of cross-site data is constructed using a sliding window method with an input window length of 16 and a prediction window length of 16. This step does not physically cut the original time series, thus preserving the continuous variation relationship between power and NWP sequences.

[0086] Step 4: Establish structural causal models for the ramp-up, plateau, and decay phases respectively, and calculate the intervention importance of each NWP feature according to formula (7). Then, use dual machine learning for biased estimation, use Bootstrap significance test and Benjamini-Hochberg false discovery rate correction to remove non-significant noise features, and then obtain the NCCIS causal importance score according to formula (8)-formula (10).

[0087] Step 5: Based on the NCCIS of the target photovoltaic power station and the reference photovoltaic power station at each stage, calculate the general causal matching degree, target specific matching degree, and comprehensive causal matching degree according to formulas (11)-(13). Remove the reference power station stage samples with a comprehensive causal matching degree lower than the preset threshold, and retain the historical samples with a consistent physical driving relationship with the target prediction scenario to obtain a large sample pool of causal homogeneity.

[0088] Step 6: Using the NWP sequence of the target prediction period as the retrieval object, retrieve candidate NWP sequences in a large causal homogeneous sample pool. Calculate the multivariate similarity between the target sequence and the candidate sequences according to formula (14), and select candidate sequences with higher multivariate similarity to form a candidate similar sequence set.

[0089] Step 7: Calculate the illumination amplitude integral or shortwave radiation integral for each candidate sequence in the candidate similar sequence set, and select the candidate sequence with the smallest integral difference from the target sequence as the most similar sequence. Use the historical measured power segment within the time window corresponding to the most similar sequence as the causal homogeneous historical response prior. For example... Figure 4 As shown, the historical response prior, NCCIS weights, stage labels, and target NWP sequence are input into the prediction model and then gated and fused with the basic output of the prediction model to obtain the final short-term photovoltaic power prediction result.

[0090] Step 8: Input the target photovoltaic power plant's NWP sequence, stage label, NCCIS weight, and historical response priors from the test set into the prediction model to obtain short-term photovoltaic power prediction values. To further verify the cross-regional adaptability, the Jilin power plant was used as the source domain, and the Hebei power plant as the target domain. Three verification methods were set up: direct source domain testing, 10% sample calibration, and 20% sample calibration.

[0091] Step 9: Calculate the error of the prediction results using the aforementioned error evaluation indicators. Compared with prediction methods that do not introduce a large, homogeneous sample pool, this embodiment can reduce the interference of historical samples with similar statistics but inconsistent physical drivers, and improve the prediction stability under complex weather and cross-regional scenarios. The prediction and scheduling indicator results are shown in Table 2.

[0092] Table 2 Prediction and scheduling indices under different methods

[0093] As can be seen from Table 2, in a specific embodiment, the short-term photovoltaic power prediction method that adopts causal homogeneous historical response priors and introduces online correction achieves lower RMSE and lower overall scheduling cost. This shows that the causal homogeneous large sample pool constructed by the present invention can transform cross-site historical information into effective historical response priors and further enhance the supporting value of prediction results for scheduling operation.

[0094] In a cross-regional validation embodiment, when the source domain model is directly migrated to the target domain, the RMSE of the prediction model after introducing causal homogeneous historical response priors and online correction is 8.232MW; after further introducing 20% ​​target domain sample calibration, the RMSE is reduced to 2.845MW.

[0095] Example 3: This application also provides a short-term photovoltaic power prediction device based on a causal homogeneous large sample pool, used to implement the method described in any of the above embodiments, such as... Figure 5 As shown, the device includes: The data preprocessing module 501 is configured to acquire historical measured power, installed capacity, timestamp data, and corresponding numerical weather forecast data of the target photovoltaic power station and multiple reference photovoltaic power stations, perform time alignment, sampling interval unification, missing value processing, and outlier processing on the data of the target photovoltaic power station and the reference photovoltaic power stations, and perform capacity normalization on the measured power according to the installed capacity of each reference photovoltaic power station, and then perform equivalent conversion according to the installed capacity of the target photovoltaic power station to obtain an equivalent power sequence at the same capacity scale as the target photovoltaic power station. The target pool construction module 502 is configured to calculate the power change slope between adjacent times based on the daily power output pattern of the target photovoltaic power station, taking the time period when the measured power of the target photovoltaic power station is greater than 0MW as the effective power output period, and determine the stage division rules of the ramp-up period, the steady period and the decay period based on the power change slope. The effective power output period of the target photovoltaic power station is marked according to the stage division rules to construct a three-stage target pool. The basic sample construction module 503 is configured to use the stage division rules of the three-stage target pool as the unified physical boundary, adopt the input window length and prediction window length consistent with the target prediction sequence, perform sliding window slicing on the complete time series data of the reference photovoltaic power station, and perform full timestamp stage pre-labeling on the complete time series to construct a cross-station basic large sample pool. The scoring calculation module 504 is configured to establish structural causal models between numerical weather forecast features and photovoltaic power for the climbing period, the stable period and the decay period, respectively. It calculates the variance of the mean of the potential response of photovoltaic power under different quantile intervention levels for each numerical weather forecast feature to obtain the intervention importance. It performs biased estimation and significance test on the intervention importance. It performs cross-site stabilization aggregation on the single-site core causal scores of the features that pass the significance test to obtain the stable core score. It performs normalization processing on the stable core score to obtain the normalized comprehensive causal importance score NCCIS. The causal homogeneous sample construction module 505 is configured to determine the stage core feature set of the target photovoltaic power station based on the NCCIS of the target photovoltaic power station at the corresponding stage, determine the cross-station general core feature set based on the NCCIS of multiple reference photovoltaic power stations at the corresponding stage, calculate the general causal matching degree of each reference photovoltaic power station relative to the cross-station general core feature set and the target specific matching degree relative to the stage core feature set of the target photovoltaic power station, determine the combined weight of the general causal matching degree and the target specific matching degree based on the causal contribution ratio of the general core features to obtain the comprehensive causal matching degree, and remove the stage samples of the reference power station with the comprehensive causal matching degree lower than the preset threshold to obtain a large causal homogeneous sample pool. The response prior calculation module 506 is configured to determine candidate numerical weather forecast sequences within the causal homogeneous large sample pool, using numerical weather forecast sequences of the target forecast period as the retrieval object; standardize the core feature dimensions of the target sequence and candidate sequences; use NCCIS as the weight to perform a weighted summation of the sequence similarity on each core feature dimension to obtain a multivariate similarity; filter the candidate similar sequence set based on the multivariate similarity; calculate the illumination amplitude integral in the candidate similar sequence set; and select the historical measured power segment within the time window corresponding to the candidate sequence with the smallest difference from the illumination amplitude integral of the target sequence as the causal homogeneous historical response prior. The short-term power prediction module 507 is configured to input the target numerical weather forecast sequence, stage labels, NCCIS weights and the causal homogeneous historical response priors into the short-term photovoltaic power prediction model, enhance the core meteorological driving features according to the NCCIS weights, constrain the feature interactions of different output stages according to the stage labels, and perform gated fusion by controlling the injection intensity of the historical response priors according to the causal homogeneity to obtain the short-term power prediction result of the target photovoltaic power station.

[0096] This application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium. When the at least one processor executes the computer program, it can implement the technical solution of the short-term photovoltaic power prediction method based on a causal homogeneous large sample pool described in the above embodiments.

[0097] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0098] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to implement the solution of this embodiment according to actual needs.

[0099] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The unit composed of the above modules can be implemented in hardware or in the form of hardware plus software functional units.

[0100] The integrated modules described above, implemented as software functional modules, can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods of the various embodiments of this application.

[0101] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor.

[0102] The memory may include high-speed RAM, and may also include non-volatile storage (NVM), such as at least one disk storage device, and may also be a USB flash drive, external hard drive, read-only memory, disk or optical disc, etc.

[0103] Buses can be Industry Standard Architecture (ISA) buses, Peripheral Component Interconnect (PCI) buses, or Extended Industry Standard Architecture (EISA) buses, etc. Buses can be categorized into address buses, data buses, control buses, etc.

[0104] The aforementioned storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0105] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor. The processor and storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and storage medium can exist as discrete components in an electronic control unit or main control device.

[0106] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A short-term photovoltaic power prediction method based on a causal homogeneous large sample pool, characterized in that, The method includes: Historical measured power, installed capacity, timestamp data, and corresponding numerical weather forecast data of the target photovoltaic power station and multiple reference photovoltaic power stations are obtained. The data of the target photovoltaic power station and the reference photovoltaic power stations are time-aligned, sampling intervals are unified, missing values ​​are handled, and outliers are handled. The measured power is normalized according to the installed capacity of each reference photovoltaic power station and then equivalently converted according to the installed capacity of the target photovoltaic power station to obtain an equivalent power sequence at the same capacity scale as the target photovoltaic power station. Based on the daily power output pattern of the target photovoltaic power station, the time period when the measured power of the target photovoltaic power station is greater than 0MW is taken as the effective power output period. The power change slope of adjacent times is calculated. Based on the power change slope, the stage division rules of the ramp-up period, the steady period and the decay period are determined. The effective power output period of the target photovoltaic power station is marked according to the stage division rules to construct a three-stage target pool. Using the phase division rules of the three-stage target pool as a unified physical boundary, and adopting the same input window length and prediction window length as the target prediction sequence, the complete time series data of the reference photovoltaic power station is sliced ​​by sliding window, and the complete time series is pre-labeled with full timestamp stages to construct a cross-station basic large sample pool. Structural causal models between numerical weather prediction features and photovoltaic power are established for the ramp-up period, the steady period, and the decay period, respectively. The variance of the mean of the potential response of photovoltaic power under different quantile intervention levels for each numerical weather prediction feature is calculated to obtain the intervention importance. The intervention importance is subjected to biased estimation and significance test. The single-site core causal scores of the features that pass the significance test are aggregated across sites to obtain the stable core score. The stable core score is normalized to obtain the normalized comprehensive causal importance score NCCIS. The target photovoltaic power station's stage core feature set is determined based on its NCCIS at the corresponding stage. A cross-station general core feature set is determined based on the NCCIS of multiple reference photovoltaic power stations at the corresponding stage. The general causal matching degree of each reference photovoltaic power station relative to the cross-station general core feature set and the target specific matching degree relative to the target photovoltaic power station's stage core feature set are calculated respectively. The combined weight of the general causal matching degree and the target specific matching degree is determined based on the causal contribution ratio of the general core features to obtain the comprehensive causal matching degree. The stage samples of the reference power stations with a comprehensive causal matching degree lower than a preset threshold are removed to obtain a large sample pool of causal homogeneity. Within the causal homogeneous large sample pool, candidate numerical weather forecast sequences are determined using numerical weather forecast sequences for the target forecast period as the retrieval object. The core feature dimensions of the target sequence and candidate sequences are standardized. The sequence similarity on each core feature dimension is weighted and summed using NCCIS as the weight to obtain multivariate similarity. Based on the multivariate similarity, a set of candidate similar sequences is selected. The illumination amplitude integral is calculated in the set of candidate similar sequences. The historical measured power segment within the time window corresponding to the candidate sequence with the smallest difference from the illumination amplitude integral of the target sequence is selected as the causal homogeneous historical response prior. The target numerical weather forecast sequence, stage labels, NCCIS weights, and the causal homogeneous historical response priors are input into the short-term photovoltaic power prediction model. The core meteorological driving features are enhanced according to the NCCIS weights, the feature interactions of different output stages are constrained according to the stage labels, and the injection intensity of the historical response priors is controlled according to the causal homogeneity to perform gated fusion, thereby obtaining the short-term power prediction results of the target photovoltaic power station.

2. The short-term photovoltaic power prediction method based on a causal homogeneous large sample pool according to claim 1, characterized in that, The equivalent power sequence of the target photovoltaic power plant at the same capacity scale is calculated using the following formula: (1) In the formula: P i ( t ) is the first i individual power station time t The measured power; P rated,i For the first i The installed capacity of each power station For the first i individual power station time t The per-unit value of power; P base The target power plant's installed capacity.

3. The short-term photovoltaic power prediction method based on a causal homogeneous large sample pool according to claim 1, characterized in that, The method involves calculating the power change slope between adjacent moments based on the daily power output pattern of the target photovoltaic power station, taking the time period when the measured power of the target photovoltaic power station is greater than 0 MW as the effective power output period, and determining the stage division rules for the ramp-up period, steady period, and decay period based on the power change slope. Specifically, this includes: The effective output period is defined as the time period during which the measured power of the target photovoltaic power station is greater than 0 MW. The slope of the power change between adjacent moments is calculated to characterize the rate of ramp-up or decay of photovoltaic power. The formula for calculating the slope of the power change is as follows: (2) In the formula: k i For the first i The slope of power change at each moment; P i , P i+1 The first i , i Measured power at time +1; Δt The data sampling time interval; The slope threshold for stage division is determined using the quantile method; the slope threshold for climbing is... k climb and attenuation slope threshold k decay These are defined as the quantiles of all candidate ramp slopes and candidate decay slopes, respectively, and the specific calculation formulas are as follows: (3) In the formula: P peak This represents the maximum power output during the effective output period of a single day; 0.85 P peak The critical power threshold for the end of the ramp-up period and the beginning of the decay period; Based on the ramp slope threshold and the decay slope threshold, a three-stage division rule is determined: power increases from the first time it is greater than 0 MW to the first time it reaches 0.85 MW. P peak The period is the ramp-up phase; power starts from 0.

85. P peak The period from the point where the power output drops to the point where it is greater than 0 MW for the last time is the decay period; the remaining period after removing the ramp-up and decay periods from the effective output period is the stable period.

4. The short-term photovoltaic power prediction method based on a causal homogeneous large sample pool according to claim 1, characterized in that, The process involves using the phase division rules of the three-stage target pool as a unified physical boundary, employing an input window length and prediction window length consistent with the target sequence to be predicted, performing sliding window slicing on the complete time series data of the reference photovoltaic power station, and pre-labeling the complete time series with full timestamp stages to construct a cross-station basic large sample pool. Specifically, this includes: Using the three-stage partitioning rule determined by the target photovoltaic power station as a unified physical boundary, and employing the same input window length and prediction window length as the target sequence to be predicted, sliding window slicing is performed on the complete time series data of the reference photovoltaic power station, and full timestamp stage pre-labeling is performed on the complete time series to obtain a cross-station basic large sample pool. D 0, the cross-site basic large sample pool D 0 is represented as: (4) In the formula: For the first i Equivalent power sequence of a reference photovoltaic power plant This is a numerical weather prediction sequence; N s The number of photovoltaic power plants is used as a reference.

5. The short-term photovoltaic power prediction method based on a causal homogeneous large sample pool according to claim 1, characterized in that, The structural causal models between numerical weather prediction features and photovoltaic power are established for the climbing, stabilizing, and decaying periods, respectively. The variance of the mean of the potential response of photovoltaic power for each numerical weather prediction feature at different quantile intervention levels is calculated to obtain the intervention importance. The intervention importance is then subjected to biased estimation and significance testing. The single-site core causal scores of features that pass the significance test are aggregated across sites to obtain a stable core score. The stable core score is then normalized to obtain the Normalized Comprehensive Causal Importance Score (NCCIS), specifically including: For the j Numerical weather forecast features X j Apply multiple quantile intervention levels x i (K) The variance of the mean potential response of photovoltaic power under different intervention levels is calculated, and this variance is used as a feature of numerical weather prediction. X j The importance of intervention The calculation formula is as follows: (5) In the formula: Indicates will X j Fixed as the number k Intervention levels; Var k Indicates to K Variance calculation of expected power response under different intervention levels; Represents the mathematical expectation; Y Indicates photovoltaic power generation capacity; Dual machine learning was used to debiased estimate the importance of the intervention, and the Bootstrap method was used for significance testing. The Benjamini-Hochberg false discovery rate was used to correct for and remove non-significant noise features. For features that passed the significance test, the single-station core causality score was calculated using the following formula: (6) In the formula: Features X i In the s The core causal score of a single photovoltaic power station; S s For the first s The set of causal significance features of a photovoltaic power station that passed the significance test; For indicator functions; For the first s Features of a photovoltaic power station X i The estimated importance of the intervention; For the first s Features of a photovoltaic power station X j The estimated importance of the intervention; Stabilized core scores are obtained by merging the individual core causal scores of multiple photovoltaic power plants. : (7) In the formula: Features X i A collection of sites with non-zero scores. ε To prevent the protection term for logarithmic infinity; if If empty, then StableCore i =0; The stable core score is normalized to obtain the normalized comprehensive causal importance score NCCIS: (8) In the formula: For the first i Numerical weather forecast features X i The normalized comprehensive causal importance score.

6. The short-term photovoltaic power prediction method based on a causal homogeneous large sample pool according to claim 5, characterized in that, Based on the NCCIS of multiple reference photovoltaic power plants at corresponding stages, a cross-site common core feature set is determined. The general causal matching degree of each reference photovoltaic power plant relative to the cross-site common core feature set and the target specific matching degree relative to the stage core feature set of the target photovoltaic power plant are calculated using the following formulas: (9) (10) In the formula, For general causal matching degree; Target-specific matching degree; and These represent the target photovoltaic power plant and the reference photovoltaic power plant at different stages. s Lower features X j Normalized comprehensive causal importance score; F core,t A cross-site universal core feature set; F target,t The core feature set for the target photovoltaic power plant stage; For indicator functions, For reference power station j In the stage t The core causal feature set below; when feature hour, Otherwise, it is 0; The comprehensive causal matching degree is calculated using the following formula: (11) (12) In the formula, For combined weights; To achieve a comprehensive causal fit.

7. The short-term photovoltaic power prediction method based on a causal homogeneous large sample pool according to claim 1, characterized in that, Within the causal homogeneous large sample pool, candidate numerical weather prediction sequences are determined using numerical weather prediction sequences for the target prediction period as the retrieval object. The core feature dimensions of the target and candidate sequences are standardized. Multivariate similarity is obtained by weighted summation of sequence similarities across each core feature dimension using NCCIS as the weight. A set of candidate similar sequences is then selected based on this multivariate similarity. Illumination amplitude integrals are calculated within the candidate similar sequence set. The historical measured power segment within the time window corresponding to the candidate sequence with the smallest difference in illumination amplitude integral from the target sequence is selected as the causal homogeneous historical response prior. Specifically, this includes: The formula for calculating multivariate similarity is as follows: (13) In the formula: For multivariate similarity; X , Y These are the target NWP sequence and the candidate NWP sequence, respectively. X i , Y i For the corresponding number i The normalized sequence of class parameters; m is the number of core NWP features involved in similarity calculation; The formula for calculating the integral of the illumination amplitude is as follows: (14) In the formula: I The integral of the illumination amplitude; G ( t ) is the first t Illumination amplitude value; Δt For time step; n This represents the total time step for the corresponding stage.

8. The short-term photovoltaic power prediction method based on a causal homogeneous large sample pool according to claim 7, characterized in that, Gated fusion is performed based on the injection intensity of the historical response prior controlled by causal homogeneity to obtain the short-term power prediction results of the target photovoltaic power station, specifically including: Gated fusion is performed based on the injection strength of the historical response prior, controlled by causal homogeneity, with the gating fusion coefficient... g The calculation formula is as follows: (15) Where: NCCIS pool Mean pooling for all core feature NCCIS scores; W is the Sigmoid activation function. g is the learnable weight matrix of the gated fusion layer; CosSim is the causal homogeneity cosine similarity between the target scene and the matching historical samples. b g For learnable bias terms of the gated fusion layer; Based on the gating fusion coefficient, the short-term power prediction result of the target photovoltaic power station is calculated using the following formula: (16) In the formula: P NWP These are the base predictions obtained by the prediction model based on the target NWP sequence; For the short-term predicted power of the target photovoltaic power plant; The causal homogeneity between the target scene and the matched historical scene is represented by the NCCIS weighted cosine similarity. When the causal homogeneity between the target scene and the historical scene is higher than a preset threshold, the gating coefficient is adjusted. g As the causal homogeneity increases, the model relies more heavily on prior causal homogeneity of historical responses to correct for NWP system bias; when the causal homogeneity is below a preset threshold, the gating coefficient... g The model is reduced, and it relies more on the underlying predictions obtained by NWP.

9. A short-term photovoltaic power prediction device based on a causal homogeneous large sample pool, characterized in that, The device includes: The data preprocessing module is configured to acquire historical measured power, installed capacity, timestamp data, and corresponding numerical weather forecast data of the target photovoltaic power station and multiple reference photovoltaic power stations. It performs time alignment, sampling interval unification, missing value processing, and outlier processing on the data of the target photovoltaic power station and the reference photovoltaic power stations. After capacity normalization of the measured power based on the installed capacity of each reference photovoltaic power station, it performs equivalent conversion according to the installed capacity of the target photovoltaic power station to obtain an equivalent power sequence at the same capacity scale as the target photovoltaic power station. The target pool construction module is configured to calculate the power change slope between adjacent times based on the daily power output pattern of the target photovoltaic power station, taking the time period when the measured power of the target photovoltaic power station is greater than 0MW as the effective power output period, and determine the stage division rules of the ramp-up period, the steady period and the decay period based on the power change slope. The effective power output period of the target photovoltaic power station is marked according to the stage division rules to construct a three-stage target pool. The basic sample construction module is configured to use the stage division rules of the three-stage target pool as the unified physical boundary, adopt the input window length and prediction window length consistent with the target prediction sequence, perform sliding window slicing on the complete time series data of the reference photovoltaic power station, and perform full timestamp stage pre-labeling on the complete time series to construct a cross-station basic large sample pool. The scoring calculation module is configured to establish structural causal models between numerical weather forecast features and photovoltaic power for the climbing period, the stable period, and the decay period, respectively. It calculates the variance of the mean of the potential response of photovoltaic power under different quantile intervention levels for each numerical weather forecast feature to obtain the intervention importance. It performs biased estimation and significance test on the intervention importance. It performs cross-site stabilization aggregation on the single-site core causal scores of the features that pass the significance test to obtain the stable core score. It performs normalization processing on the stable core score to obtain the normalized comprehensive causal importance score NCCIS. The causal homogeneous sample construction module is configured to determine the stage core feature set of the target photovoltaic power station based on the NCCIS of the target photovoltaic power station at the corresponding stage, determine the cross-station general core feature set based on the NCCIS of multiple reference photovoltaic power stations at the corresponding stage, calculate the general causal matching degree of each reference photovoltaic power station relative to the cross-station general core feature set and the target specific matching degree relative to the stage core feature set of the target photovoltaic power station, determine the combined weight of the general causal matching degree and the target specific matching degree based on the causal contribution ratio of the general core features to obtain the comprehensive causal matching degree, and remove the stage samples of the reference power station with the comprehensive causal matching degree lower than the preset threshold to obtain a large causal homogeneous sample pool. The response prior calculation module is configured to determine candidate numerical weather forecast sequences within the causal homogeneous large sample pool, using numerical weather forecast sequences of the target forecast period as the retrieval object; standardize the core feature dimensions of the target sequence and candidate sequences; use NCCIS as the weight to perform a weighted summation of the sequence similarity on each core feature dimension to obtain a multivariate similarity; filter the candidate similar sequence set based on the multivariate similarity; calculate the illumination amplitude integral in the candidate similar sequence set; and select the historical measured power segment within the time window corresponding to the candidate sequence with the smallest difference from the illumination amplitude integral of the target sequence as the causal homogeneous historical response prior. The short-term power prediction module is configured to input the target numerical weather forecast sequence, stage labels, NCCIS weights, and the causal homogeneous historical response priors into the short-term photovoltaic power prediction model. It enhances the core meteorological driving features according to the NCCIS weights, constrains the feature interactions of different output stages according to the stage labels, and performs gated fusion by controlling the injection intensity of the historical response priors according to the causal homogeneity to obtain the short-term power prediction results of the target photovoltaic power station.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the short-term photovoltaic power prediction method based on a causal homogeneous large sample pool as described in any one of claims 1-8.

Images