Photovoltaic power prediction method and system

Abstract

The invention relates to the technical field of photovoltaic power prediction, and discloses a photovoltaic power prediction method and system, and the method comprises the steps: obtaining the historical operation data of each photovoltaic station, carrying out the abnormal value elimination, expanding the sample data through a generative adversarial network, constructing a first training sample set, and carrying out the variable dimension reduction, screening principal component factors influencing the photovoltaic power to construct a second training sample set; calculating similar days by using the second training sample set, and screening and sorting; establishing a deep learning framework fusing the long short-term memory network, the maximum temperature prediction model and a parameter optimization algorithm, and based on a preset photovoltaic power prediction precision evaluation index, performing model training by taking the number of days of similar days and the weight as independent variables to establish a power prediction model; based on preset reanalysis data and weather forecast data, photovoltaic power prediction is carried out by using the trained power prediction model, precision evaluation and dynamic optimization of model parameters are carried out, and the photovoltaic power prediction precision and adaptability to different scenes are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power prediction, in particular to a photovoltaic power prediction method and system. BACKGROUND

[0002] High-precision photovoltaic power prediction technology plays an important supporting role in optimizing power generation plans, improving power generation efficiency for new energy enterprises, and achieving power supply and demand balance management and ensuring safe and stable operation of power grids for power grid operators.

[0003] Existing photovoltaic power prediction methods mainly rely on current daily weather forecast data and historical power sequences to construct prediction models. Such models usually use time series analysis methods to achieve power prediction by mining time series characteristics in historical data. However, due to the complexity and uncertainty of weather conditions, and the time-varying nature of the mapping relationship between weather elements and photovoltaic power at different time scales, the model generalization ability is insufficient, and the prediction accuracy significantly decreases under complex weather conditions.

[0004] Some advanced prediction techniques use similar day analysis methods to filter historical similar scenario data by using weather condition similarity, effectively enhancing the model's adaptability to complex weather conditions. However, this method still has the following key problems in practical application: 1. Insufficient training samples, short construction period of new energy stations, limited accumulation of historical operation data, especially lack of sample data under extreme weather conditions, leading to insufficient model training, making it difficult to accurately capture power variation characteristics under complex weather conditions.

[0005] 2. Low computational efficiency, numerous weather element variables (such as temperature, humidity, light intensity, cloud cover, etc.), and high complexity of similarity calculation in high-dimensional feature space. With the increase in the number of similar days, the computational overhead increases exponentially, seriously affecting the real-time prediction.

[0006] 3. Parameter optimization problem, existing similar day filtering and weight distribution methods mostly rely on expert experience or fixed parameter settings, lacking adaptive optimization mechanism. The optimal parameters of different geographical regions and different scale stations differ significantly, and fixed parameter configuration leads to poor model portability, making it difficult to meet diverse application requirements. SUMMARY

[0007] Therefore, the present application provides a photovoltaic power prediction method and system to solve the problems of insufficient samples, low computational efficiency, and strong subjectivity of parameters in traditional photovoltaic power prediction models.

[0008] In a first aspect, the present application provides a photovoltaic power prediction method, comprising: The historical operating data of each photovoltaic station is obtained to remove outliers, and the sample data is expanded through the generative adversarial network to construct the first training sample set; Performing variable dimension reduction on the first training sample set by principal component analysis, screening principal component factors affecting photovoltaic power, and constructing a second training sample set; Calculate the similarity between the historical date and the current date using the second training sample set, and filter and sort similar days based on the similarity; A deep learning framework was established that integrated a long-short-term memory network, a maximum temperature prediction model, and a parameter optimization algorithm. Based on preset photovoltaic power prediction accuracy evaluation indicators, the model was trained using the number of similar days and weights as independent variables to establish a power prediction model. Based on the preset reanalysis data and meteorological forecast data, the trained power prediction model is used to predict the photovoltaic power of the site, and the accuracy evaluation and dynamic optimization of model parameters are carried out.

[0009] The photovoltaic power prediction method provided by the embodiment of the present invention expands samples and eliminates outliers to improve data quality by generating an adversarial network, optimizes computing efficiency by using principal component analysis to reduce dimensionality and screen core features, uses similarity calculation to intelligently screen similar days and realizes adaptive weight adjustment through a parameter optimization algorithm, constructs a deep learning framework that integrates a long-short-term memory network and a maximum temperature prediction model to enhance the capture of nonlinear relationships, and dynamically optimizes model parameters by combining reanalysis data with a rolling prediction mechanism, thereby significantly improving the accuracy of photovoltaic power prediction, providing reliable data support for power grid dispatching, helping power generation companies optimize their operating strategies, and enhancing the adaptability of the model to different scenarios.

[0010] In an optional embodiment, the acquisition of historical operating data of each photovoltaic station to remove outliers, and the expansion of sample data by generating an adversarial network to construct a first training sample set include: Obtain data from each PV station within a preset historical time period, including historical power generation, meteorological data affecting PV power, and station scheduling information; Based on the station dispatch information, timestamps are added based on the data labeling method to remove outliers caused by artificial power restrictions in the historical power generation of photovoltaic stations. On this basis, a second outlier removal is performed based on the time window moving average method; A generative adversarial network is constructed, and the historical power generation power and meteorological data of each photovoltaic station are input into the generator for sample expansion to form the first training sample set.

[0011] The embodiment of the present invention obtains multi-source information such as historical power generation power, meteorological data and scheduling information of photovoltaic stations to provide comprehensive basic data for model training; uses the data labeling method to add timestamps and eliminate anomalies caused by artificial power restrictions, and combines the time window moving average method for secondary elimination to doubly ensure data reliability and reduce noise interference; uses a generative adversarial network to expand samples of historical power generation power and meteorological data, effectively solving the problems of insufficient data accumulation of new energy stations and scarcity of extreme scenario samples. The expanded first training sample set can more comprehensively characterize the variation law of photovoltaic power, laying a high-quality data foundation for subsequent model training.

[0012] In an optional embodiment, performing variable dimension reduction on the first training sample set by principal component analysis, screening principal component factors affecting photovoltaic power, and constructing the second training sample set includes: Normalizing the power generation and meteorological data in the first training sample set; Use principal component analysis to reduce the dimension of the normalized sample data; A threshold is set according to the contribution rate, the comprehensive index of the principal component factor is screened, and the second training sample set is constructed in combination with the historical power generation at the corresponding moment.

[0013] The embodiment of the present invention normalizes the power generation and meteorological data in the first training sample set to eliminate the influence of data of different dimensions and ensure the accuracy of principal component analysis; uses principal component analysis to reduce the dimension of the normalized data and converts high-dimensional correlated variables into low-dimensional uncorrelated principal component factors, thereby significantly reducing the data dimension while retaining key information and improving computational efficiency; screens principal component factors according to the contribution rate threshold and constructs the second training sample set in combination with historical power generation, which can not only focus on the core factors affecting photovoltaic power but also avoid feature redundancy, provide concise and effective feature vectors for subsequent similar day analysis and model training, and significantly optimize data processing efficiency and model training performance. In an optional embodiment, based on a preset photovoltaic power prediction accuracy evaluation index, the number of similar days and weights are used as independent variables, a deep learning framework integrating a long short-term memory network, a maximum temperature prediction model, and a parameter optimization algorithm is established and model training is performed to establish a power prediction model, including: Set accuracy evaluation indicators, including accuracy and root mean square error; Using the second training sample set as input, the PV power during the forecast period as output, accuracy and root mean square error as evaluation indicators, and the number of similar days and weights as independent variables to be optimized, a deep learning framework was established that integrates a long-short-term memory network, a maximum temperature prediction model, and a preset parameter optimization algorithm. Use particle swarm optimization or PEST algorithm to optimize the number of similar days and weight parameters; The maximum temperature prediction model outputs the maximum temperature in the forecast period as reinforcement training data for the long short-term memory network model; Using a long short-term memory network and a preset loss function as the optimization target, a preset relationship between the principal component factor, generated power, predicted temperature, and predicted power is established; The deep learning framework is trained in an end-to-end joint training manner, and the validation set is used to perform parameter optimization with accuracy and root mean square error as evaluation indicators to establish a power prediction model. The model training scheme of the embodiment of the present invention significantly improves prediction accuracy and generalization capability through multi-technology integration and adaptive optimization: accuracy and root mean square error are used as quantitative indicators to ensure that model accuracy can be evaluated; the long short-term memory network (LSTM), maximum temperature prediction (MTP) model and particle swarm optimization (PSO) / PEST parameter optimization algorithm are integrated to construct an LMP deep learning framework, in which the PSO / PEST algorithm automatically optimizes the number of similar days and weights to avoid manual experience bias. The maximum temperature of the forecast period output by the MTP model is used as LSTM reinforcement training data to enhance the ability to capture the impact of temperature on power. LSTM establishes a nonlinear mapping relationship between principal component factors, generated power, predicted temperature and predicted power with the loss function as the target; through end-to-end joint training and validation set parameter optimization, the model accurately fits the complex relationship between meteorological factors and power changes. The final power prediction model has both high robustness and adaptive tuning capabilities, which can effectively meet the power prediction needs in different scenarios. In an optional embodiment, the method of performing site photovoltaic power forecasting based on preset reanalysis data and meteorological forecast data using a trained power forecast model, and performing accuracy evaluation and dynamic optimization of model parameters includes: Obtain meteorological data sets formed by fusion of multi-source data as reanalysis data, including: meteorological historical data and future forecast data; Using the meteorological historical data of the reanalysis data as input, a weather forecast model dedicated to photovoltaic sites is established by optimizing the parameterization scheme of the numerical model; Inputting the future prediction data of the reanalysis data into the weather forecast model and outputting the weather forecast result of the photovoltaic station; Based on the meteorological forecast results of the photovoltaic station, the process of repeatedly constructing the first and second training sample sets for prediction and screening and sorting similar days is repeated. The historical power data is input into the power prediction model to perform rolling prediction of photovoltaic power and dynamic optimization of the model.

[0014] The meteorological coordination and rolling forecasting solution provided by the embodiment of the present invention achieves continuous improvement in forecast accuracy through multi-source data fusion and dynamic optimization mechanism, and uses multi-source fused reanalysis data (including historical and future forecast data) to provide high-reliability input for meteorological modeling. By optimizing the cloud microphysics, boundary layer and other parameterization schemes of the WRF model, a meteorological forecast model adapted to the area where the photovoltaic station is located is established to improve the accuracy of local meteorological forecasts. Future forecast data of the reanalysis data is input into the model to generate accurate station meteorological forecast results. Based on the meteorological forecast repeated data preprocessing, feature dimensionality reduction and similar day screening process, a forecast sample set is constructed and combined with historical power data for rolling forecast. At the same time, the training set is updated after each forecast result is fused with historical data, driving the LMP deep learning framework to dynamically optimize parameters, so that the model can adapt to changes in meteorological conditions in real time, continuously optimize the forecast accuracy, and provide decision support for station power forecasting with both timeliness and accuracy. In an optional embodiment, the performing of rolling photovoltaic power prediction and model dynamic optimization includes: After each rolling forecast of photovoltaic power, the new meteorological data and power forecast results are integrated with the historical second training sample set for iterative optimization of the deep learning framework to support subsequent rolling forecasts of photovoltaic power.

[0015] The embodiment of the present invention forms a dynamically updated training data pool by integrating the new meteorological data and power forecast results obtained from each rolling forecast with the historical second training sample set, providing the deep learning framework with training samples that are closer to real-time scenarios, enabling the model to continuously learn the latest meteorological change patterns and power fluctuation characteristics, and effectively solving the problem of attenuation of model prediction accuracy caused by time-varying meteorological conditions. Through this iterative optimization mechanism, the model's adaptability to emerging meteorological patterns and prediction accuracy are continuously improved, ensuring that the photovoltaic power rolling forecast always maintains high reliability in long-term operation, and providing more accurate real-time data support for site power generation plan formulation and grid scheduling. In a second aspect, the present invention provides a photovoltaic power prediction system, the system comprising: The first training sample set construction module is used to obtain the historical operating data of each photovoltaic station to remove outliers, and expand the sample data through the generative adversarial network to construct the first training sample set; A second training sample set construction module is used to perform variable dimension reduction on the first training sample set through principal component analysis, screen principal component factors that affect photovoltaic power, and construct a second training sample set; A similar day screening and sorting module is used to calculate the similarity between historical dates and the current day using the second training sample set, and to screen and sort similar days based on the similarity; The power prediction model construction module is used to establish a power prediction model based on the preset photovoltaic power prediction accuracy evaluation index, using the number of similar days and weights as independent variables, and to build a deep learning framework that integrates the long-short-term memory network, the maximum temperature prediction model, and the parameter optimization algorithm, and perform model training; The model dynamic optimization module is used to predict the photovoltaic power of the site based on the preset reanalysis data and meteorological forecast data using the trained power prediction model, and to perform accuracy evaluation and dynamic optimization of model parameters.

[0016] In a third aspect, the present invention provides a computer device comprising: a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the photovoltaic power prediction method of the first aspect or any corresponding embodiment thereof by executing the computer instructions.

[0017] In a fourth aspect, the present invention provides a computer-readable storage medium having computer instructions stored thereon, the computer instructions being used to enable a computer to execute the photovoltaic power prediction method of the first aspect or any corresponding embodiment thereof.

[0018] In a fifth aspect, the present invention provides a computer program product, comprising computer instructions for causing a computer to execute the photovoltaic power prediction method of the first aspect or any corresponding embodiment thereof. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0020] Figure 1 is a schematic flow chart of a photovoltaic power prediction method according to an embodiment of the present invention; Figure 2 is a schematic diagram of end-to-end joint training of a photovoltaic power prediction model according to an embodiment of the present invention; Figure 3 is a structural block diagram of a photovoltaic power prediction system according to an embodiment of the present invention; Figure 4 Schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0021] To make the purpose, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative efforts shall fall within the scope of protection of the present invention.

[0022] In order to overcome the shortcomings of traditional photovoltaic power prediction models such as insufficient samples, low computational efficiency, and strong parameter subjectivity, a photovoltaic power prediction method is provided in this embodiment. Figure 1 FIG. 1 is a flow chart of a photovoltaic power prediction method according to an embodiment of the present invention. Figure 1 As shown, the process includes the following steps: Step S1: Obtain historical operating data of each photovoltaic station to remove outliers, and expand sample data through a generative adversarial network to construct a first training sample set. Specifically, step S1 includes the following steps: S11, obtaining data from each photovoltaic station within a preset historical time period, including the historical power generation of the photovoltaic station, meteorological data affecting photovoltaic power, and station scheduling information; S12, based on the station scheduling information, add timestamps based on the data labeling method to eliminate outliers caused by artificial power restrictions in the historical power generation of the photovoltaic station, and then perform a secondary outlier elimination based on the time window moving average method; The embodiment of the present invention uses the time window moving average method to eliminate outliers twice, which is expressed by the following formula:

[0023] Where x t represents the photovoltaic power at time t; T represents the length of the time window; Indicates the average photovoltaic power within the preset historical time period; Indicates the lower limit of power deviation.

[0024] S13, constructing a generative adversarial network, inputting the historical power generation and meteorological data of each photovoltaic station into the generator for sample expansion to form a first training sample set.

[0025] In one embodiment, taking a 300MW photovoltaic station as an example, its historical operation data covers January 2022 to June 2024. The specific processing flow is as follows: 1. Multi-source data acquisition: Collect the station's historical power generation (15-minute resolution), meteorological data (including irradiance, temperature, humidity, wind speed, etc.), and scheduling logs (such as planned maintenance and grid power restrictions). For example, in one scenario, due to tight grid load, the station was subject to power restrictions from 2:00 PM to 4:00 PM on July 15, 2023. Power generation remained at 60% of rated power, significantly deviating from the normal operating curve.

[0026] 2. Double elimination of outliers: Step 1 (labeling method): Mark the power restriction period from 14:00 to 16:00 on July 15, 2023 according to the scheduling log, add a timestamp and remove the corresponding power data.

[0027] Step 2 (Time Window Moving Average): Set a time window of T = 24 hours (i.e., 1 day) and calculate the historical power mean and lower deviation limit for the station. If the power at a certain moment falls below the lower deviation limit and does not fall within the marked power curtailment period, it is considered a noise outlier and is removed. For example, a sudden power drop at 8:00 AM on March 5, 2024, due to a sensor failure, was detected using the moving average method and was removed twice.

[0028] 3. Generative Adversarial Network Sample Expansion: A generative adversarial network (GAN) model is constructed, generating data through adversarial training between the generator and discriminator. In one embodiment, for example, the generator inputs historical power (12,000 records) and meteorological data (irradiance, temperature, etc.) from 2022 to 2024 for the station. The generator outputs simulated samples that align with the distribution of the real data (e.g., data for scarce scenarios such as extreme high temperatures and heavy rainfall). After expansion, the sample size increases to 28,000 records. The generated data includes the power data and the corresponding meteorological data. For example, power data is generated for the "three consecutive days of high temperatures + strong winds" scenario, which did not occur in 2023, to enhance the model's adaptability to extreme weather conditions.

[0029] This embodiment of the present invention uses a scheduling log labeling method to accurately locate outliers in artificial power curtailment, avoiding model training bias caused by grid dispatch intervention. A time-window moving average method further filters random noise such as sensor failures and communication interruptions, ensuring that training data more closely reflects actual power generation patterns. A generative adversarial network model learns the probability distribution of historical data to generate power samples for rare meteorological scenarios (such as extreme heat and rare typhoons), filling the sample gap caused by the short operating time of new energy stations. This combination of precise denoising and intelligent augmentation fundamentally addresses the core issues of photovoltaic power forecasting, namely, limited data, poor quality, and a lack of scenarios, providing a high-quality training foundation for subsequent modeling.

[0030] Step S2, performing variable dimension reduction on the first training sample set by principal component analysis, screening the principal component factors affecting photovoltaic power, and constructing a second training sample set. Specifically, step S2 includes the following steps: Step S21: normalize the power generation and meteorological data in the first training sample set; specifically, the following steps are performed:

[0031] Where x new Represents the normalized element value (power value, meteorological element value); x i represents the value of the element at time i; x min 、x max The normalization process ensures that meteorological factors of different dimensions (such as irradiance and temperature) are in the same numerical range, thus avoiding weight bias caused by differences in feature scales during model training.

[0032] Step S22, use principal component analysis to reduce the dimension of the normalized sample data; the principal component analysis used in the embodiment of the present invention is a statistical method that uses orthogonal transformation to transform a group of original random vectors with possibly correlated components into new random vectors with uncorrelated components, forming a new principal component comprehensive index, and replacing the original more variable indicators with a few comprehensive indicators. For example, the definition: x 1, x 2,…, x p is the original variable indicator, z 1, z 2,…, z m (m <p)为新变量指标,则:

[0033] Step S23 , setting a threshold according to the contribution rate, screening the comprehensive index of the principal component factor, and constructing a second training sample set in combination with the historical power generation at the corresponding moment.

[0034] In one embodiment, for example, 12-dimensional meteorological factors (irradiance, temperature, humidity, wind speed, wind direction, air pressure, cloud cover, etc.) and power generation are input, the covariance matrix is ​​calculated, and the eigenvalues ​​and eigenvectors are extracted. The contribution rates of the first three principal components are as follows: Principal component 1 (PC1): Contribution rate 45%, mainly reflecting the coupling effect of irradiance and temperature; Principal component 2 (PC2): Contribution rate 28%, mainly reflecting the correlation between wind speed and air pressure; Principal component 3 (PC3): Contribution rate 15%, mainly reflecting the fluctuations of cloud cover and humidity.

[0035] The contribution rate threshold is set to 80% (45% + 28% + 15% = 88% ≥ 80%), and the 12-dimensional features are finally reduced to 3-dimensional principal component factors. The 3-dimensional principal component factors are combined with the power generation power (normalized) at the corresponding moment to form a new sample.

[0036] The embodiment of the present invention automatically screens principal component factors based on contribution rates, avoiding manual subjective feature selection. Through the progressive processing of "normalization-dimensionality reduction-screening", it achieves the optimal balance between data efficiency and model performance. It is particularly suitable for photovoltaic power prediction scenarios with complex meteorological factors and high data dimensions, and provides efficient feature representation for subsequent similar day analysis and deep learning modeling.

[0037] Step S3: Calculate the similarity between the historical date and the current date using the second training sample set, and filter and sort similar days based on the similarity.

[0038] In the embodiment of the present invention, similar days refer to several historical dates with weather conditions similar to the current date. Similarity is generally used as the evaluation criterion. The similarity can be calculated using methods such as Euclidean distance, cosine similarity, and dynamic time warping (DTW). In the present invention, the principal component factors of the second sample set after dimensionality reduction are used as input. Taking the Euclidean distance method as an example, the specific method is:

[0039] Where d(*) represents the Euclidean distance between two sample points; Z i and Z j Represents two points in the n-dimensional feature space. For time series, each point can include multiple feature factors ( z i1 , z i2 ,..., z in ; z j1 , z j2 ,..., z jn ). Furthermore, the Euclidean distance is used to calculate the similarity between the current day and the historical date, and all historical dates are sorted in ascending order of distance, and the top n similar days are selected. n can be reasonably set according to the specific scenario requirements.

[0040] The embodiment of the present invention uses the feature vectors after dimensionality reduction of the principal component factors to effectively characterize the coupling relationship of core meteorological elements. Since the computational effort is reduced after dimensionality reduction, the time required to screen similar days at the station can be significantly reduced, meeting the real-time requirements of short-term power forecasting. The similar day screening mechanism of the embodiment of the present invention significantly improves computational efficiency while ensuring meteorological matching accuracy through the combination of "dimensionality reduction features + Euclidean distance," providing high-quality training samples for subsequent deep learning frameworks. This is particularly suitable for photovoltaic stations with variable meteorological conditions and complex factor coupling, effectively enhancing the robustness and adaptability of the power forecasting model.

[0041] Step S4: Establish a deep learning framework that integrates a long-short-term memory network, a maximum temperature prediction model, and a parameter optimization algorithm. Based on the preset photovoltaic power prediction accuracy evaluation index, the model is trained with the number of similar days and weights as independent variables to establish a power prediction model. Specifically, step S4 includes the following steps: S41, set the accuracy evaluation indicators, including accuracy and root mean square error; accuracy (ACC) and root mean square error (RMSE) are specifically expressed as:

[0042]

[0043] Where, C represents the maximum startup capacity of the photovoltaic station on the assessment day; n represents the number of hours in the forecast period; P M,i Indicates the actual power at the time to be predicted; P P,i Indicates the predicted power at the time to be predicted.

[0044] S42, using the second training sample set as input, the photovoltaic power in the period to be predicted as output, the accuracy and root mean square error as evaluation indicators, and the number of similar days and weights as independent variables to be optimized, establish a deep learning framework that integrates the long short-term memory network, the maximum temperature prediction model and the preset parameter optimization algorithm.

[0045] It combines the time series feature capture capability of LSTM, the environmental factor prediction of the temperature prediction model, and the feature dimensionality reduction of principal component analysis to solve the problem of nonlinear prediction of photovoltaic power affected by multiple factors such as weather and equipment status.

[0046] S43, using a particle swarm algorithm or a PEST algorithm to optimize the number of similar days and weight parameters.

[0047] The present invention dynamically optimizes the number of similar days and weights through the PSO or PEST algorithm, so that the model can adapt to the power fluctuation laws in different seasons and weather conditions, and improve the prediction stability in special scenarios.

[0048] S44, outputting the maximum temperature of the forecast period through the maximum temperature prediction model as reinforcement training data for the long short-term memory network model.

[0049] In embodiments of the present invention, the maximum temperature prediction model may rely on a high-resolution WRF model, combined with real-time radar echoes and satellite cloud imagery, to capture sudden impacts of local convective weather on temperature (e.g., cooling caused by thunderstorms), output an hourly temperature forecast sequence, and take the maximum value as the daily maximum temperature forecast. Alternatively, an ensemble learning model may be employed, integrating large-scale circulation characteristics (e.g., the location of the subtropical high pressure) from a numerical forecast model (e.g., the ECMWF medium-term forecast) with regional climate characteristics learned from a statistical model (e.g., the probability of high temperatures in summer) to generate a probability distribution of daily maximum temperatures for the next three days. The peak of the probability density function is taken as the predicted value. This is for illustrative purposes only and is not intended to be limiting.

[0050] The power generation efficiency of photovoltaic modules is strongly correlated with temperature (typically, silicon-based module efficiency decreases by approximately 0.3% to 0.5% for every 1°C increase in temperature). Temperature variations (such as the periodicity of daily maximum temperatures) exhibit distinct temporal patterns and are highly coupled to diurnal fluctuations in photovoltaic power (e.g., early morning increases and late evening decreases). Using temperature as input helps the LSTM memory unit more accurately capture the dynamic mapping between power and temperature, learning long-term co-variation patterns. This can reduce trend errors caused by seasonal temperature differences, especially when dealing with cross-day and cross-season forecasting tasks.

[0051] S45, using a long short-term memory network with a preset loss function as an optimization target, establishes a preset relationship among the principal component factors, the generated power, the predicted temperature and the predicted power.

[0052] S46, adopts end-to-end joint training to train the deep learning framework, and uses the validation set to optimize parameters with accuracy and root mean square error as evaluation indicators to establish a power prediction model.

[0053] The embodiment of the present invention uses a long-short-term memory network with a preset loss function as the optimization target to establish a preset relationship between principal component factors, power generation power, predicted temperature and predicted power, which can effectively integrate multi-source heterogeneous data. The principal component factors realize data dimensionality reduction and eliminate redundancy, and combine historical power generation power and predicted temperature to comprehensively cover power influencing factors; the preset loss function accurately guides model parameter optimization, and the LSTM gating mechanism captures long-term and short-term dependencies, improving the model's adaptability to complex scenarios; at the same time, the design enhances the generalization ability and interpretability of the model, achieving a leap from data fitting to law abstraction, and interacting with modules such as temperature prediction to form an end-to-end closed-loop optimization, significantly improving the accuracy and reliability of photovoltaic power prediction, and providing strong decision-making support for power grid scheduling, power station operation and maintenance, etc.

[0054] The schematic diagram of end-to-end joint training of the photovoltaic power prediction model based on the LMP deep learning framework that integrates the long short-term memory network (LSTM), maximum temperature prediction (MTP) model and PEST parameter optimization algorithm based on the second training sample set is as follows: Figure 2 This embodiment of the present invention uses the LSTM, MTP model, and parameter optimization algorithm as an overall framework, optimizing all parameters simultaneously to avoid error accumulation in step-by-step training. End-to-end training eliminates the need for manual parameter adjustment or step-by-step data processing. The framework automatically integrates data preprocessing, feature extraction, model training, and parameter optimization, making it suitable for real-time prediction scenarios in large-scale photovoltaic stations.

[0055] Step S5, based on the preset reanalysis data and weather forecast data, uses the trained power prediction model to predict the photovoltaic power of the site, and performs accuracy evaluation and dynamic optimization of model parameters. Specifically, step S5 includes the following steps: S51, obtains the meteorological data set formed by the fusion of multi-source data as reanalysis data, including: meteorological historical data and future forecast data; among them, reanalysis data refers to the meteorological data set formed by the fusion of multi-source data such as model data and observation data, which provides meteorological historical return data and future forecast data. Common ones include ERA5, GFS and ECMWF. The reanalysis data formed by the fusion of multi-source data integrates meteorological historical and future forecast data, retaining historical laws and incorporating future trends, providing a rich information basis for accurate prediction.

[0056] S52, using the meteorological historical data of the reanalysis data as input, establishes a meteorological prediction model dedicated to the photovoltaic station by optimizing the parameterization scheme of the numerical model; its numerical model can be the WRF model, and its parameterization scheme includes cloud microphysical process scheme, boundary layer scheme, land surface process scheme, etc. It is necessary to adjust the parameterization scheme according to regional characteristics and with reference to the measured meteorological data, and construct a meteorological model dedicated to the photovoltaic station based on the optimal numerical model parameterization scheme based on historical data. This can specifically solve the problem of insufficient accuracy of the general meteorological model when applied to a specific station, and fit the actual meteorological environment of the station.

[0057] S53, input the future prediction data of the reanalysis data into the meteorological prediction model, and output the meteorological forecast results of the photovoltaic station; input the future prediction data into the dedicated meteorological model, and the output meteorological forecast results can accurately reflect the meteorological changes of the station and provide a reliable meteorological basis for power prediction.

[0058] S54, based on the weather forecast results of the photovoltaic station, repeat the process of constructing the first training sample set and the second training sample set for prediction and screening and sorting similar days, and combine the historical power data to input the power prediction model to perform rolling prediction of photovoltaic power and dynamic optimization of the model. The embodiment of the present invention repeats the sample set construction and screening and sorting of similar days based on the weather forecast results, and combines the historical power data to perform rolling prediction and dynamic optimization of the model, so that the model can adapt to meteorological conditions and power fluctuations in real time, and continuously improve the accuracy of the prediction. Finally, through precision evaluation and dynamic optimization of model parameters, it is ensured that the prediction model continues to maintain high accuracy, meets the precision requirements of photovoltaic power stations in scenarios such as grid dispatching, power trading, operation and maintenance management, effectively reduces the abandonment rate, and improves the economic benefits of the power station and the stability of the power system.

[0059] It should be noted that after each rolling PV power forecast, the present invention integrates the latest meteorological data and power forecast results with the second historical training sample set for iterative optimization of the deep learning framework to support subsequent rolling PV power forecasts. By integrating the latest meteorological data and power forecast results with the second historical training sample set and using them for iterative optimization of the deep learning framework to support subsequent rolling forecasts, this design establishes a dynamic closed loop of "prediction-feedback-optimization," bringing significant advantages in terms of data updating, model evolution, and application value. First, at the data level, the real-time integration of new meteorological data (such as sudden rainstorms, extreme high temperatures, etc.) and forecast results can quickly supplement the sample set, fill in special scenarios not covered by historical data, make the training data more in line with current reality, and solve the problem of insufficient sample timeliness caused by changing meteorological conditions; secondly, at the model level, iterative optimization allows the deep learning framework to dynamically adapt to environmental fluctuations. For example, when cloudy weather occurs frequently, the model can automatically adjust the weights of features such as cloud thickness and light intensity through new data to avoid the decline of forecast performance due to environmental changes; finally, at the application value level, the continuously optimized model significantly improves the long-term stability of rolling forecasts, reduces the accumulation of forecast errors, helps power stations to accurately quote in power transactions, optimize resource allocation in grid dispatching, reduce economic losses and operation and maintenance costs caused by forecast deviations, and ultimately realize the intelligent, adaptive and efficient operation of photovoltaic systems.

[0060] This embodiment also provides a photovoltaic power prediction system for implementing the above-mentioned embodiments and preferred implementations. Details already described will not be repeated here. As used below, the term "module" may refer to a combination of software and / or hardware that implements a predetermined function. While the systems described in the following embodiments are preferably implemented in software, implementation in hardware, or a combination of software and hardware, is also possible and contemplated.

[0061] This embodiment provides a photovoltaic power prediction system. Figure 3 Shown, including: A first training sample set construction module 31 is used to obtain historical operating data of each photovoltaic station to remove outliers, and expand the sample data through a generative adversarial network to construct a first training sample set; A second training sample set construction module 32 is configured to perform variable dimension reduction on the first training sample set by principal component analysis, screen principal component factors that affect photovoltaic power, and construct a second training sample set; A similar day screening and sorting module 33 is used to calculate the similarity between the historical date and the current day using the second training sample set, and to screen and sort the similar days based on the similarity; A power prediction model building module 34 is used to establish a deep learning framework integrating a long-short-term memory network, a maximum temperature prediction model, and a parameter optimization algorithm based on a preset photovoltaic power prediction accuracy evaluation index and using the number of similar days and weights as independent variables, and to perform model training to establish a power prediction model; The model dynamic optimization module 35 is used to predict the photovoltaic power of the site based on the preset reanalysis data and meteorological forecast data using the trained power prediction model, and to perform accuracy evaluation and dynamic optimization of model parameters.

[0062] In some optional implementations, the first training sample set construction module 31 includes: A historical data acquisition unit is used to obtain data from each photovoltaic station within a preset historical time period, including the historical power generation of the photovoltaic station, meteorological data affecting photovoltaic power, and station scheduling information; The outlier elimination unit is used to remove outliers caused by artificial power restrictions in the historical power generation of photovoltaic stations based on the station scheduling information and the data labeling method, and then perform secondary outlier elimination based on the time window moving average method. The first training sample set construction unit is used to construct a generative adversarial network, input the historical power generation power and meteorological data of each photovoltaic station into the generator for sample expansion, and form a first training sample set.

[0063] In some optional implementations, the second training sample set construction module 32 includes: a normalization processing unit, configured to perform normalization processing on the power generation and meteorological data in the first training sample set; A data dimension reduction unit is used to reduce the variable dimension of the normalized sample data using principal component analysis; The second training sample set construction unit is used to set a threshold according to the contribution rate, screen the comprehensive index of the main component factor, and construct the second training sample set in combination with the historical power generation at the corresponding moment.

[0064] In some optional implementations, the power prediction model building module 34 includes: Evaluation index setting unit, used to set accuracy evaluation indicators, including accuracy and root mean square error; a deep learning framework construction unit, configured to use the second training sample set as input, the photovoltaic power in the period to be predicted as output, accuracy and root mean square error as evaluation indicators, and the number of similar days and weights as independent variables to be optimized, to establish a deep learning framework integrating a long short-term memory network, a maximum temperature prediction model, and a preset parameter optimization algorithm; Among them, the particle swarm algorithm or PEST algorithm is used to optimize the number of similar days and weight parameters; the maximum temperature of the forecast period is output by the maximum temperature prediction model as reinforcement training data for the long-short-term memory network model; the long-short-term memory network is used to establish a preset relationship between the principal component factor, power generation, predicted temperature and predicted power with a preset loss function as the optimization target; The model training unit is used to train the deep learning framework in an end-to-end joint training manner, and to optimize parameters using a validation set with accuracy and root mean square error as evaluation indicators to establish a power prediction model.

[0065] In an optional embodiment, the model dynamic optimization module 35 includes: The reanalysis data acquisition unit is used to obtain the meteorological data set formed by the fusion of multi-source data as reanalysis data, including: meteorological historical data and future forecast data; The meteorological forecast model building unit is used to use the meteorological historical data of the reanalysis data as input and establish a meteorological forecast model dedicated to the photovoltaic site by optimizing the parameterization scheme of the numerical model; A weather forecast result acquisition unit, configured to input future forecast data of the reanalysis data into the weather forecast model and output a weather forecast result for the photovoltaic station; The model dynamic optimization unit is used to repeatedly construct the first and second training sample sets for prediction and screen and sort similar days based on the meteorological forecast results of the photovoltaic station, and input the historical power data into the power prediction model to perform rolling prediction of photovoltaic power and dynamic optimization of the model.

[0066] In an optional embodiment, after each rolling forecast of photovoltaic power, the model dynamic optimization unit integrates the new meteorological data and power forecast results with the historical second training sample set for iterative optimization of the deep learning framework to support subsequent rolling forecasts of photovoltaic power.

[0067] The further functional description of each of the above modules and units is the same as that of the above corresponding embodiments and will not be repeated here.

[0068] The photovoltaic power prediction system in this embodiment is presented in the form of functional units, where the units refer to ASIC (Application Specific Integrated Circuit) circuits, processors and memories that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0069] The embodiment of the present invention also provides a computer device having the above Figure 3 The photovoltaic power prediction system shown.

[0070] See also Figure 4 , Figure 4 is a schematic diagram of the structure of a computer device provided by an optional embodiment of the present invention, such as Figure 4 As shown, the computer device includes: one or more processors 10, memory 20, and interfaces for connecting various components, including high-speed interfaces and low-speed interfaces. Various components utilize different buses to communicate with each other and can be installed on a common mainboard or installed in other ways as needed. The processor can process the instructions executed in the computer device, including instructions stored in the memory or on the memory to display the graphical information of the GUI on an external input / output device (such as, a display device coupled to the interface). In some optional embodiments, if necessary, multiple processors and / or multiple buses can be used together with multiple memories and multiple memories. Equally, multiple computer devices can be connected, and each device provides part of the necessary operations (for example, as a server array, a group of blade servers, or a multi-processor system). Figure 4 A processor 10 is taken as an example.

[0071] The processor 10 may be a central processing unit, a network processor, or a combination thereof. The processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit, a programmable logic device, or a combination thereof. The programmable logic device may be a complex programmable logic device, a field programmable gate array, a general purpose array logic, or any combination thereof.

[0072] The memory 20 stores instructions that can be executed by at least one processor 10, so as to enable at least one processor 10 to execute the method shown in the above embodiment.

[0073] The memory 20 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and application programs required for at least one function; the data storage area may store data created based on the use of the computer device, etc. In addition, the memory 20 may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some optional embodiments, the memory 20 may optionally include a memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0074] The memory 20 may include a volatile memory, such as a random access memory; the memory may also include a non-volatile memory, such as a flash memory, a hard disk or a solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0075] The computer device further includes a communication interface 30 for the computer device to communicate with other devices or a communication network.

[0076] The embodiment of the present invention also provides a computer-readable storage medium. The above-mentioned method according to the embodiment of the present invention can be implemented in hardware, firmware, or implemented as a computer code that can be recorded in a storage medium, or implemented as a computer code that is originally stored in a remote storage medium or a non-temporary machine-readable storage medium and downloaded through a network and will be stored in a local storage medium, so that the method described herein can be stored in such software processing on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only storage memory, a random access memory, a flash memory, a hard disk or a solid-state drive, etc.; further, the storage medium can also include a combination of the above-mentioned types of memory. It can be understood that a computer, a processor, a microprocessor controller or programmable hardware includes a storage component that can store or receive software or computer code. When the software or computer code is accessed and executed by a computer, a processor or hardware, the method shown in the above embodiment is implemented.

[0077] A portion of the present invention may be applied as a computer program product, such as a computer program instruction, which, when executed by a computer, can call or provide the method and / or technical solution according to the present invention through the operation of the computer. Those skilled in the art should understand that the form in which the computer program instruction exists in a computer-readable medium includes, but is not limited to, a source file, an executable file, an installation package file, etc. Accordingly, the way in which the computer program instruction is executed by the computer includes, but is not limited to: the computer directly executes the instruction, or the computer compiles the instruction and then executes the corresponding compiled program, or the computer reads and executes the instruction, or the computer reads and installs the instruction and then executes the corresponding installed program. Here, the computer-readable medium may be any available computer-readable storage medium or communication medium that can be accessed by the computer.

[0078] Although the embodiments of the present invention have been described with reference to the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention. Such modifications and variations are all within the scope defined by the appended claims.

Claims

1. A photovoltaic power prediction method, characterized in that: include: The historical operating data of each photovoltaic station is obtained to remove outliers, and the sample data is expanded through the generative adversarial network to construct the first training sample set; Performing variable dimension reduction on the first training sample set by principal component analysis, screening principal component factors affecting photovoltaic power, and constructing a second training sample set; Calculate the similarity between the historical date and the current date using the second training sample set, and filter and sort similar days based on the similarity; A deep learning framework was established that integrated a long-short-term memory network, a maximum temperature prediction model, and a parameter optimization algorithm. Based on preset photovoltaic power prediction accuracy evaluation indicators, the model was trained using the number of similar days and weights as independent variables to establish a power prediction model. Based on the preset reanalysis data and meteorological forecast data, the trained power prediction model is used to predict the photovoltaic power of the site, and the accuracy evaluation and dynamic optimization of model parameters are carried out.

2. The method according to claim 1, characterized in that The method of obtaining historical operating data of each photovoltaic station to remove outliers and expanding sample data by generating an adversarial network to construct a first training sample set includes: Obtain data from each PV station within a preset historical time period, including historical power generation, meteorological data affecting PV power, and station scheduling information; Based on the station dispatch information, timestamps are added based on the data labeling method to remove outliers caused by artificial power restrictions in the historical power generation of photovoltaic stations. On this basis, a second outlier removal is performed based on the time window moving average method; A generative adversarial network is constructed, and the historical power generation power and meteorological data of each photovoltaic station are input into the generator for sample expansion to form the first training sample set.

3. The method according to claim 1, characterized in that The step of performing variable dimension reduction on the first training sample set by principal component analysis, screening principal component factors affecting photovoltaic power, and constructing a second training sample set includes: Normalizing the power generation and meteorological data in the first training sample set; Use principal component analysis to reduce the dimension of the normalized sample data; A threshold is set according to the contribution rate, the comprehensive index of the principal component factor is screened, and the second training sample set is constructed in combination with the historical power generation at the corresponding moment.

4. The method according to claim 1 or 3, characterized in that Based on the preset photovoltaic power prediction accuracy evaluation index, the number of similar days and weights are used as independent variables to establish a deep learning framework that integrates a long-short-term memory network, a maximum temperature prediction model, and a parameter optimization algorithm, and perform model training to establish a power prediction model, including: Set accuracy evaluation indicators, including accuracy and root mean square error; Using the second training sample set as input, the PV power during the forecast period as output, accuracy and root mean square error as evaluation indicators, and the number of similar days and weights as independent variables to be optimized, a deep learning framework was established that integrates a long-short-term memory network, a maximum temperature prediction model, and a preset parameter optimization algorithm. Use particle swarm optimization or PEST algorithm to optimize the number of similar days and weight parameters; The maximum temperature prediction model outputs the maximum temperature in the forecast period as reinforcement training data for the long short-term memory network model; Using a long short-term memory network and a preset loss function as the optimization target, a preset relationship between the principal component factor, generated power, predicted temperature, and predicted power is established; The deep learning framework is trained in an end-to-end joint training manner, and the validation set is used to perform parameter optimization with accuracy and root mean square error as evaluation indicators to establish a power prediction model.

5. The method according to claim 1, wherein The method of predicting photovoltaic power at a site using a trained power prediction model based on preset reanalysis data and meteorological forecast data, and performing accuracy evaluation and dynamic optimization of model parameters includes: Obtain meteorological data sets formed by fusion of multi-source data as reanalysis data, including: meteorological historical data and future forecast data; Using the meteorological historical data of the reanalysis data as input, a weather forecast model dedicated to photovoltaic sites is established by optimizing the parameterization scheme of the numerical model; Inputting the future prediction data of the reanalysis data into the weather forecast model and outputting the weather forecast result of the photovoltaic station; Based on the meteorological forecast results of the photovoltaic station, the process of repeatedly constructing the first and second training sample sets for prediction and screening and sorting similar days is repeated. The historical power data is input into the power prediction model to perform rolling prediction of photovoltaic power and dynamic optimization of the model.

6. The method according to claim 5, characterized in that The photovoltaic power rolling forecast and model dynamic optimization include: After each rolling forecast of photovoltaic power, the new meteorological data and power forecast results are integrated with the historical second training sample set for iterative optimization of the deep learning framework to support subsequent rolling forecasts of photovoltaic power.

7. A photovoltaic power prediction system, characterized in that: include: The first training sample set construction module is used to obtain the historical operating data of each photovoltaic station to remove outliers, and expand the sample data through the generative adversarial network to construct the first training sample set; A second training sample set construction module is used to perform variable dimension reduction on the first training sample set through principal component analysis, screen principal component factors that affect photovoltaic power, and construct a second training sample set; A similar day screening and sorting module is used to calculate the similarity between historical dates and the current day using the second training sample set, and to screen and sort similar days based on the similarity; The power prediction model construction module is used to establish a power prediction model based on the preset photovoltaic power prediction accuracy evaluation index, using the number of similar days and weights as independent variables, and to build a deep learning framework that integrates the long-short-term memory network, the maximum temperature prediction model, and the parameter optimization algorithm, and perform model training; The model dynamic optimization module is used to predict the photovoltaic power of the site based on the preset reanalysis data and meteorological forecast data using the trained power prediction model, and to perform accuracy evaluation and dynamic optimization of model parameters.

8. A computer device, characterized in that: include: A memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the photovoltaic power prediction method according to any one of claims 1 to 6 by executing the computer instructions.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a computer to execute the photovoltaic power prediction method according to any one of claims 1 to 6.

10. A computer program product, characterized in that The method comprises computer instructions for causing a computer to execute the photovoltaic power prediction method according to any one of claims 1 to 6.

Images