Photovoltaic intra-day power supply guarantee capability prediction method and device

Abstract

The invention discloses a photovoltaic intra-day power supply guarantee capability prediction method and device, and relates to the technical field of power system optimization scheduling, and the method comprises the steps: processing the multi-dimensional input parameters of a research region based on a convolutional neural network-bidirectional long-short-term memory network model when the photovoltaic intra-day power supply guarantee capability is evaluated, and carrying out the prediction of the photovoltaic intra-day power supply guarantee capability. The convolutional neural network-bidirectional long short-term memory network model is trained based on a multi-objective loss function, and the multi-objective loss function comprehensively considers penalty terms such as prediction precision, down-repair degree, positive deviation recognition and response speed, so that a more accurate photovoltaic power generation guarantee output lower limit can be obtained. The prediction accuracy and reliability of the photovoltaic power generation power in the aspect of the minimum reliable output are improved, and the accuracy and reliability of evaluating the photovoltaic intra-day power supply guarantee capability are further improved, so that the scheduling risk caused by overestimation of the photovoltaic output is reduced, the problem of unreasonable scheduling plan in deterministic prediction is solved, and the prediction efficiency is improved. The stability of the power system is improved.

Description

Technical Field

[0001] This application relates to the field of power system optimization and dispatching technology, and in particular to a method and device for predicting the daily power supply guarantee capacity of photovoltaic systems. Background Technology

[0002] The rapid development of photovoltaic (PV) power generation and its increasing share in the electricity supply place higher demands on the stable operation of the power system. PV power generation is affected by various factors, exhibiting volatility and intermittency. This poses significant challenges to operational aspects of the power system, such as reserve capacity analysis, maintenance planning, and unit commitments. The power grid requires more accurate and reliable PV power generation forecasting methods to address its volatility and uncertainty, ensuring the stability and reliability of the power supply. However, existing PV power generation forecasting methods suffer from insufficient deterministic prediction, affecting the intraday power supply guarantee capability of PV systems and consequently impacting the accuracy of power dispatching plans, ultimately failing to guarantee the stability of the power supply system. Summary of the Invention

[0003] The purpose of this application is to provide a method and apparatus for predicting the daily power supply guarantee capability of photovoltaic systems, which can improve the accuracy and reliability of predicting the daily power supply guarantee capability of photovoltaic systems, enhance the accuracy of power dispatching plans, and thus ensure the stability of the power system.

[0004] To achieve the above objectives, this application provides the following solution:

[0005] Firstly, this application provides a method for predicting the intraday power supply guarantee capability of photovoltaic systems, including:

[0006] This study predicts the minimum guaranteed output of photovoltaic (PV) power generation, periods of low PV output, and the probability of PV power supply guarantee for future periods in the research area. The minimum guaranteed output refers to the predicted minimum reliable PV power output for future periods. Periods of low output refer to the time when the first power difference exceeds the supplementary capacity of the PV backup units. The first power difference is the difference between the PV power output on a classic clear-sky day and the actual PV power output. The probability of PV power supply guarantee refers to the probability that the predicted PV power output for future periods can meet specific power supply needs. Based on the minimum guaranteed output, periods of low PV output, and the probability of PV power supply guarantee, the intraday PV power supply guarantee capability for future periods is assessed.

[0007] The process of predicting the guaranteed lower limit of photovoltaic power generation output specifically includes:

[0008] A multidimensional input parameter is obtained for the study area, including current time information, gridded irradiance data for future time periods, gridded wind speed data for future time periods, and historical actual photovoltaic power generation. Based on the multidimensional input parameter, a trained guaranteed output lower limit prediction model is used to determine the guaranteed output lower limit for photovoltaic power generation in the future time period. The guaranteed output lower limit prediction model includes a convolutional neural network and a bidirectional long short-term memory network. The guaranteed output lower limit prediction model is trained based on a multi-objective loss function, which is a function constructed based on a prediction accuracy penalty term, a prediction positive deviation penalty term, a smoothness penalty term, and a response speed penalty term. The prediction accuracy penalty term measures the prediction accuracy of the guaranteed output lower limit prediction model. The prediction positive deviation penalty term measures the magnitude of the prediction positive deviation, which is a variable determined based on the relationship between the guaranteed output lower limit and the actual photovoltaic power generation. The smoothness penalty term measures the smoothness of the guaranteed output lower limit over time. The response speed penalty term measures the sensitivity of the guaranteed output lower limit prediction model to changes in the actual photovoltaic power generation.

[0009] In a second aspect, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the photovoltaic intraday power supply guarantee capability prediction method described in the first aspect above.

[0010] According to the specific embodiments provided in this application, this application has the following technical effects:

[0011] This application provides a method and apparatus for predicting the daily power supply guarantee capability of photovoltaic power generation. When assessing the daily power supply guarantee capability of photovoltaic power generation, this method processes multi-dimensional input parameters of the study area using a convolutional neural network-bidirectional long short-term memory network model. The convolutional neural network-bidirectional long short-term memory network model is trained based on a multi-objective loss function, which comprehensively considers penalty terms such as prediction accuracy, downward revision degree, positive deviation identification, and response speed. This allows for a more accurate lower limit of photovoltaic power generation guarantee output (the minimum reliable value of photovoltaic output), forming a conservative estimate of photovoltaic output. This improves the accuracy and reliability of photovoltaic power generation prediction in terms of minimum reliable output, increases power supply reliability and the proportion of guaranteed power, thereby improving the accuracy and reliability of the daily power supply guarantee capability of photovoltaic power generation in the study area in the future. It also reduces the dispatch risk caused by overestimating photovoltaic output, solves the problem of unreasonable dispatch plans in deterministic prediction, and enhances the stability of the power system. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating the prediction of the lower limit of photovoltaic power generation output in a method for predicting the intraday power supply guarantee capacity of photovoltaic power in Embodiment 1 of this application.

[0014] Figure 2 This is a schematic diagram of the provincial photovoltaic power output curve in Embodiment 1 of this application;

[0015] Figure 3 This is a schematic diagram of the predicted output of solar power under sunny conditions in Embodiment 1 of this application;

[0016] Figure 4 This is a schematic diagram of the photovoltaic power output prediction results for cloudy and sunny days in Embodiment 1 of this application;

[0017] Figure 5 This is a schematic diagram of the photovoltaic power output prediction results for rainy and sunny days in Embodiment 1 of this application;

[0018] Figure 6 This is a flowchart illustrating the prediction of low-output periods of photovoltaic power generation in a method for predicting the intraday power supply guarantee capacity of photovoltaic power in Embodiment 1 of this application.

[0019] Figure 7 This is a conceptual diagram of the low-output process prediction research scheme in Embodiment 1 of this application;

[0020] Figure 8 This is a schematic diagram comparing the actual photovoltaic power generation in Embodiment 1 of this application with the power output of a classic clear-sky model day;

[0021] Figure 9 This is a schematic diagram comparing the processes of merging adjacent low-output units when the backup unit capacity is 15% of the total photovoltaic installed capacity of the province in Embodiment 1 of this application.

[0022] Figure 10 This is a flowchart illustrating the hyperparameter optimization process of the low-output period prediction model based on grid search optimization in Embodiment 1 of this application.

[0023] Figure 11 This is a flowchart illustrating the photovoltaic power supply guarantee probability prediction in a photovoltaic intraday power supply guarantee capability prediction method according to Embodiment 1 of this application;

[0024] Figure 12This is a schematic diagram of the daily provincial photovoltaic power generation survival function in Embodiment 1 of this application;

[0025] Figure 13 This is a schematic diagram of the cumulative distribution function of the integral of the probability prediction result at a certain moment in Embodiment 1 of this application;

[0026] Figure 14 The following is a schematic diagram of the cumulative distribution function set at multiple time scales for the photovoltaic intraday power supply guarantee probability prediction results in Example 1 of this application.

[0027] Figure 15 This is a schematic diagram of zero-probability deletion in Embodiment 1 of this application;

[0028] Figure 16 This is a schematic diagram of the 95% confidence interval results for January 19, 2021, in Embodiment 1 of this application;

[0029] Figure 17 This is a schematic diagram of the power prediction probability density function graph at 12:30 on January 19, 2021, in Embodiment 1 of this application;

[0030] Figure 18 This is a schematic diagram of the probability distribution function graph after adjusting the prediction results at different confidence levels at 12:30 on January 19, 2021, in Embodiment 1 of this application;

[0031] Figure 19 This is a schematic diagram of the cumulative power distribution function after zero-probability deletion at 12:30 on January 19, 2021, in Embodiment 1 of this application;

[0032] Figure 20 This is a schematic diagram of the photovoltaic intraday power supply guarantee probability prediction results in Embodiment 1 of this application;

[0033] Figure 21 This is a schematic diagram of the probability prediction results for the 0.2 pu power supply demand section on January 1, 2022, in Embodiment 1 of this application;

[0034] Figure 22 This is a graph showing the relationship between the Brier score and accuracy percentage of the test set and the power supply demand profile in Embodiment 1 of this application. Detailed Implementation

[0035] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0036] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0037] Example 1

[0038] Research has revealed that existing photovoltaic power prediction technologies suffer from insufficient deterministic prediction: grid dispatching agencies rely heavily on deterministic prediction results to formulate power generation plans, but these results cannot quantify the uncertainty of errors, which can easily lead to unreasonable proportions of new energy sources being included. This seriously affects the accuracy and reliability of the prediction of photovoltaic intraday power supply guarantee capacity, resulting in overly conservative or aggressive dispatching plans, which in turn affects the effective utilization of photovoltaic resources and the stable supply of the power system.

[0039] In response, this embodiment provides a method for predicting the daily power supply guarantee capability of photovoltaic power generation, including: (1) predicting the lower limit of photovoltaic power generation guarantee output, the period of low photovoltaic power generation output, and the probability of photovoltaic power supply guarantee in the future period of the study area, wherein the lower limit of photovoltaic power generation guarantee output refers to the predicted minimum reliable power of photovoltaic power generation in the future period, the period of low output refers to the time when the first power difference is greater than the supplementary capacity of the photovoltaic backup unit, the first power difference refers to the difference between the photovoltaic power generation power on a classic clear day and the actual photovoltaic power generation power, and the probability of photovoltaic power supply guarantee refers to the probability that the predicted photovoltaic power generation power in the future period can meet specific power supply needs; (2) evaluating the daily power supply guarantee capability of photovoltaic power generation in the future period based on the lower limit of photovoltaic power generation guarantee output, the period of low photovoltaic power generation output, and the probability of photovoltaic power supply guarantee.

[0040] like Figure 1 As shown, the prediction process for the guaranteed lower limit of photovoltaic power generation output specifically includes:

[0041] S1, Obtain multi-dimensional input parameters for the study area, wherein the multi-dimensional input parameters include current time information, regional gridded irradiance data for future periods, and historical actual photovoltaic power generation; S2, Based on the multi-dimensional input parameters, determine the guaranteed output limit for photovoltaic power generation in future periods using a trained guaranteed output limit prediction model, wherein the guaranteed output limit prediction model includes Convolutional Neural Networks (CNN) and Bi-directional Long Short-Term Memory (BiLSTM).

[0042] The following describes the specific process of predicting the lower limit of photovoltaic power supply guarantee output during the day.

[0043] Step 1: Input Feature Selection. Select multi-dimensional input features, including time features, gridded predicted irradiance, and historical actual photovoltaic power.

[0044] Step 2: Data Preprocessing. In the data preprocessing step, outlier filtering, missing data interpolation, and data normalization were performed to optimize model training efficiency and accuracy.

[0045] Step 3: Predict guaranteed photovoltaic output (i.e., the lower limit of guaranteed photovoltaic power output) based on CNN-BiLSTM. In CNN-BiLSTM, the preprocessed data is analyzed in depth. First, the CNN part is responsible for extracting the spatial feature information of the provincial gridded predicted irradiance, revealing the spatial distribution information of photovoltaic power plants. Second, the data flows to the BiLSTM layer, which focuses on mining the temporal patterns of guaranteed output and capturing the features that evolve over time. Finally, a fully connected layer integrates spatial and temporal information to accurately output the temporal features of guaranteed photovoltaic output.

[0046] Step 4: Multi-objective Loss Function and Model Optimization. In deterministic prediction, mean squared error is often used as the loss function for training neural network parameters. The main goal is to reduce the total difference between the predicted and actual values. However, this method does not fully consider the impact of error magnitude on grid dispatch security. Overestimating photovoltaic output may lead to insufficient grid dispatch, failing to meet peak-hour power demand and affecting power supply. Therefore, a loss function is needed to correct the prediction results to improve the accuracy and reliability of grid dispatch. This embodiment, based on the constructed CNN-BiLSTM model, integrates multiple constraints, including prediction accuracy, prediction power correction strength, and the model's fast response capability. By optimizing the loss function through a multi-objective function, a photovoltaic guaranteed output with the goal of ensuring safety is obtained. The optimal configuration of model parameters is determined by stopping training after more than 50 iterations.

[0047] Step 5: Output photovoltaic power output to ensure safety. After the spatial distribution information and time series features are extracted by the CNN-BiLSTM model, the data is passed to the output layer to generate accurate photovoltaic power output prediction values.

[0048] The following calculation example verifies the effectiveness of the photovoltaic intraday power supply guarantee output prediction method proposed in this embodiment.

[0049] 1.1 Definition of photovoltaic power output guarantee. Figure 2 This document presents a comparison chart of the predicted and actual photovoltaic power curves for a certain province within a day. The chart shows that the predicted power is higher than the actual power during certain periods. This situation is referred to as "positive prediction deviation" in this embodiment. The magnitude of the positive prediction deviation reflects the overestimation of the actual power by the predicted power.

[0050] When predicting power Greater than the actual power P i When, the predicted positive deviation y iThe absolute error between the two is considered; when the predicted power is less than or equal to the actual power, there is no overestimation. In this case, the positive prediction deviation y... i Set to 0MW. Predict positive deviation y i Definition:

[0051] Large positive deviations in forecasts can lead to misjudgments by dispatchers, negatively impacting power supply and system safety. This embodiment aims to apply artificial intelligence technology to predict the daily guaranteed output of provincial photovoltaic power, ensuring that the predicted curve is lower than the actual power curve, thus forming a conservative estimate of photovoltaic output. In this embodiment, this curve is defined as the guaranteed photovoltaic output curve. Figure 2 As shown by the red curve.

[0052] 1.2 Provincial-level photovoltaic daily guaranteed output prediction model based on CNN and BiLSTM networks. The number of model layers was optimized using the ablation experiment method. The model structure includes two CNN layers and two BiLSTM layers. The CNN extracts the spatial features of the input gridded predicted irradiance and gridded predicted wind speed data. Through convolution operations, it extracts the irradiance differences between different grid points and the irradiance variation information within a specific time period, reflecting the spatial changes in irradiance data. The BiLSTM layer processes time-series data. Through its unique forward and backward memory capabilities, it captures historical time features and the long-term dependence of historical photovoltaic actual power. The BiLSTM optimizes the capture of time information at both ends of the sequence by combining its unique forward and backward LSTM units. The forward LSTM unit is responsible for capturing historical information and extracting patterns useful for future prediction from past time-series data. Simultaneously, the backward LSTM unit processes future information, enabling the model to utilize the latest data up to the prediction time, improving the prediction accuracy and response speed when predicting guaranteed output.

[0053] 1.3 Multi-objective optimization of a custom loss function. To achieve accurate prediction of guaranteed photovoltaic power output, the predicted value of the guaranteed power output needs to simultaneously satisfy multiple constraints ( To ensure the predicted output value): (1) The goal is to ensure that the predicted output is as accurate as possible, so as to achieve a sufficiently high accuracy in the predicted output. (2) The predicted output value is required to be less than the actual power value, thereby ensuring that the output value is below the actual power value. (3) The predicted output value is guaranteed to be non-negative. (4) Accelerate the response speed of prediction models.

[0054] Based on the above constraints, a multi-objective loss function is designed, and the value of the multi-objective loss function is minimized by introducing multiple penalty terms. A multi-objective optimization strategy is adopted to finely adjust the loss function and achieve optimal prediction performance. The expression of the multi-loss objective function is: Loss=J1+J2+J3+J4 (2).

[0055] The individual penalty items are defined as follows:

[0056] (1) J1 is the prediction accuracy penalty term:

[0057] The prediction accuracy penalty term J1 aims to narrow the gap between the model's predicted value and the expected value, ensuring the overall accuracy of the predicted value of the guaranteed output.

[0058] (2) J2 is the penalty term for the degree of positive deviation in prediction:

[0059] The goal of J2 design is to penalize the model when the predicted guaranteed output is greater than the actual power. The loss value of J2 is adjusted by the positive deviation coefficient λ, where λ > 0 to ensure that the degree of positive deviation J2 increases as the predicted excess increases, thus ensuring that the guaranteed output is below the actual output.

[0060] (3) J3 is the prediction smoothness penalty term:

[0061] The goal of J3 design is to limit excessive fluctuations in forecast values ​​and make forecast results smoother.

[0062] (4) J4 is a fast response penalty item:

[0063] In the formula: γ is the adjustment parameter for the fast response penalty term; t is the current time; t-1 is the previous time. The goal of J4 is to enhance the model's sensitivity to rapid changes in actual power. This penalty term uses the change in actual power between the current and previous time points as the benchmark for applying the penalty. Simultaneously, the penalty intensity is adjusted by introducing the parameter γ. If the change in actual power between the current and previous time points is large, J4 will adjust the model parameters to stimulate the model's fast response capability, enabling the model to react quickly to changes in the data.

[0064] The optimal prediction model parameters are found by minimizing the multiple loss objective function, Loss. This process takes place in the model hypothesis space, where the model parameters are adjusted to minimize the total loss function, thereby enabling the discovery of the actual power time-series correlation patterns.

[0065] 1.4 Evaluation Indicators for Prediction Results: To evaluate the prediction results of the photovoltaic power supply guarantee curve, two types of evaluation indicators are defined. The first type measures the closeness between the guaranteed output curve and the actual power curve, introducing normalized mean absolute error, normalized mean square error, and coefficient of determination. The second type mainly reflects the power supply guarantee capability, including the guarantee rate and the guaranteed power ratio.

[0066] (1) Coverage Ratio (GR):

[0067]

[0068] The guarantee rate represents the proportion of samples where the guaranteed output is lower than the actual power output. It reflects the proportion of positive deviations and the reliability of the power supply. A higher guarantee rate means a higher probability that the actual power output is higher than the guaranteed output, a lower proportion of positive deviations, and higher power supply reliability; conversely, a lower probability that the actual power output is higher than the guaranteed output means a higher proportion of positive deviations and lower power supply reliability.

[0069] (2) Guaranteed Energy Ratio (GER):

[0070] The guaranteed power ratio represents the area under the guaranteed output curve and the actual power curve. This indicator reflects the accuracy of the guaranteed output prediction and, to some extent, the power supply guarantee capability. Specifically, it assesses whether the guaranteed output prediction meets the demand, using actual power and electricity as a benchmark (actual photovoltaic power as the required output power). A higher guaranteed power ratio indicates higher prediction accuracy and stronger power supply guarantee capability; a lower guaranteed power ratio indicates lower prediction accuracy and a risk of insufficient power supply. It is important to emphasize that assessing the guaranteed power ratio requires high reliability. This means the guarantee rate must be at a high level. Insufficient guarantee rate will lead to an excessively high positive deviation ratio, increasing the risk that the predicted output of provincial photovoltaic power plants will fall short of the actual output. In this case, evaluating their power supply guarantee capability is meaningless.

[0071] 1.5 Case Study Verification: The data comes from 195 photovoltaic (PV) power stations in a northern province of my country, covering the provincial PV output and daily guaranteed output forecasts from January 2021 to May 2022. This embodiment also uses gridded irradiance data provided by the CAMS Solar Radiation Service Database. Given the significant intermittent nature of PV output, this embodiment selects data from 05:00 to 20:30 daily, analyzing it at 15-minute sampling intervals. The prediction results are as follows... Figures 3-5 As shown, the overall guarantee rate of photovoltaic power supply guarantee output prediction reached 98.16%, and the guaranteed power volume accounted for 93.61%, which well completed the power supply guarantee output prediction task with the goal of ensuring safety.

[0072] This embodiment introduces a multi-objective loss function encompassing four penalty terms: prediction accuracy, prediction positive bias identification, prediction positive bias degree, and response speed, to mine the temporal characteristics of prediction power positive bias. The proposed method can reduce positive bias while maintaining high overall prediction accuracy, exhibiting higher accuracy compared to other conventional loss functions.

[0073] Research has found that factors affecting the daily power supply guarantee capability of photovoltaic systems also include periods of low photovoltaic power generation. However, related technologies still suffer from inaccurate identification of these low-output periods. Therefore, solutions such as... Figure 6 The prediction process for low-output periods of photovoltaic power generation in this embodiment specifically includes: Sa, acquiring target input data for the study area, wherein the target input data includes current time information, regional gridded irradiance data for future periods, the guaranteed output limit of photovoltaic power generation for future periods, and the photovoltaic power generation power on a classic clear-sky day, wherein the guaranteed output limit of photovoltaic power generation for future periods is a variable determined using the photovoltaic power generation prediction method described in Embodiment 1; Sb, determining the low-output periods of photovoltaic power generation for future periods using a trained low-output period prediction model based on the target input data, wherein the low-output period prediction model is a machine learning model.

[0074] This embodiment can more accurately identify periods of low power output, providing a basis for dispatching to take measures in advance to address the imbalance between power supply and demand.

[0075] The following describes the specific process of the photovoltaic low-output process (i.e., low-output period) prediction method.

[0076] Step 1: Define a low-output process. Step 2: Encode sample features based on actual photovoltaic power output data, and construct historical training samples considering NWP meteorological factors related to output levels. In data preprocessing, merging low-output periods avoids fragmentation of prediction results due to data fluctuations, and identifying short-term power fluctuations through anomaly data distinguishes them from true low-output processes, improving the quality of training samples. Step 3: Optimize the low-output identification model parameters based on grid search, and calculate the pattern recognition result (0 or 1) and probability of the low-output process. Step 4: Propose corresponding evaluation indicators to assess the prediction results of the low-output process.

[0077] like Figure 7 The proposed low-output process prediction research scheme includes:

[0078] 2.1 Definition of low-output process: Figure 8 This figure shows a comparison between the actual power generation of photovoltaic systems and the power output of a classic clear-sky model day. As can be seen from the figure, the actual power P at some times t... tThe power output P of the day is lower than that of the classic clear-sky model. C (t), the difference is higher than the backup unit supplementary capacity P. backup The set of all times is called the period T during which the low-output process occurs. low The formula is described as follows:

[0079] T low ={t|P C (t)-P(t)>P backup ,t∈T} (10)

[0080] P backup =θ×P total (11)

[0081] In the formula: P C P(t) represents the power output on a classic clear-sky day, P(t) represents the actual output power, and P represents the backup unit replenishment capacity. backup Generally, the total installed photovoltaic capacity P of the entire province is taken as the standard. total The low output threshold θ ranges from 10% to 20%.

[0082] 2.2 Low-Output Process Sample Construction: Low-output sample construction refers to encoding the photovoltaic power output level at each time point of the day into a binary sequence based on the actual photovoltaic power output data, to represent the power output status at each time point, for subsequent model training and testing. This encoding method can effectively identify low-output periods in the photovoltaic power generation system, providing reliable label data for the low-output identification model.

[0083] (1) Actual low output coding

[0084] The low-output process reflects the start and end times and duration of the gap where the backup unit capacity cannot meet the actual output below the expected level. For each time point t, if the power output P on a classic clear-sky day... C (t) The difference between the actual output power Pt and the backup unit's supplementary capacity Pt is lower than the backup unit's supplementary capacity Pt. backup If the output is low (coded as "1"), it is recorded as low output; otherwise, it is recorded as normal or high output (coded as "0").

[0085]

[0086] (2) Conventional “NWP-Predicted Power-Low Output Calibration” method

[0087] The conventional method "NWP-Predicted Power-Low Output Calibration" serves as a control group, using the photovoltaic power generation predicted by a photovoltaic power prediction model with NWP as input to replace the actual measured power for low output calibration. In practical applications, actual photovoltaic output power data is unavailable; therefore, the predicted power of the photovoltaic power prediction model with NWP as input can be used as an approximation for calibrating the low output process.

[0088]

[0089] In the formula: The "NWP-Predicted Power-Low Output Calibration" method utilizes the predicted power of the NWP model at time point t. The calibration results are used as an approximation of the actual power P(t) for low-output processes.

[0090] 2.3 Historical Sample Preprocessing: Merging adjacent low-output periods on the same day aims to improve the continuity and accuracy of low-output process prediction, and avoid intermittent low-output coding prediction results caused by data fluctuations, thus preventing the omission of individual periods. Figure 9 This is a comparison chart showing the merging of adjacent low-output processes when the backup unit capacity is 15% of the total photovoltaic installed capacity in the province. To distinguish short-term low-output fluctuations from true low-output processes, the minimum duration of a low-output process is defined as exceeding 15 minutes to eliminate short-term low-output fluctuations. Subsequently, irradiance and power data are subjected to min-max normalization, transforming data of different dimensions and ranges to the [0,1] interval. This avoids the influence of certain features on low-output identification parameters due to their large original scale, improving the efficiency and performance of model training.

[0091] In terms of time feature processing, time features are transformed by sin and cos to preserve periodic information while normalizing the time features.

[0092]

[0093] In the formula: T sin (t), T cos (t) represents the normalized time feature; t and T represent the original time feature and corresponding time period. The time features include year, month, day, hour, and minute. Due to the intermittent nature of photovoltaic power output, the photovoltaic power data is extracted from 05:00am to 20:30pm daily, and the sample contains 15.5 hours per day.

[0094] 2.4 Low-Output Period Prediction Model Based on Grid Search Optimization: To optimize the low-output period prediction model, this example uses a grid search and 5-fold cross-validation method to select the optimal hyperparameter combination, with accuracy as the evaluation metric. The optimization flowchart is shown below. Figure 10 As shown, grid search finds the optimal parameter combination by systematically traversing a specified set of parameters, while 5-fold cross-validation evaluates the model's performance by training it multiple times and finally selects the parameter combination with the best average performance.

[0095] 2.5 Evaluation Indicators for Low-Output Period Forecasts: Based on the judgment rules for low-output period forecast results, evaluation indicators for low-output period forecasts are constructed, including recall rate, accuracy rate, false alarm rate, missed alarm rate, key success index, and fairness skill score.

[0096] (1) Recall: Recall refers to the model's ability to correctly identify low-output processes, i.e., the ratio of processes predicted as low-output processes that actually are to the total number of processes that are actually low-output processes. The index ranges from 0 to 1. The closer the value is to 1, the higher the recall, the more effective the model is in identifying and capturing low-output processes, and the better the model performance. (2) Accuracy: Accuracy refers to the model's ability to correctly identify low-output processes and normal processes, i.e., the ratio of processes predicted as low-output processes that actually are to the total number of processes predicted as low-output processes. The index ranges from 0 to 1. The closer the value is to 1, the higher the accuracy, the more accurate the model is in distinguishing between low-output events and normal events, and the better the model performance. (3) False Alarm Rate: False alarm rate refers to the rate at which the model incorrectly identifies a normal process as a low-output process, i.e., the ratio of processes predicted as low-output processes but actually being normal to the total number of samples predicted as low-output processes. The index ranges from 0 to 1. The lower the false alarm rate, the fewer false alarms the model predicts, and the higher the prediction quality. (4) False alarm rate: The false alarm rate refers to the model's ability to fail to identify low-output processes, that is, the ratio of processes predicted as normal but actually being low-output processes to processes that are actually low-output processes. The false alarm rate ranges from 0 to 1, and the lower the value, the more effective the model is in identifying and capturing all potential low-output processes.

[0097] (5) Key Success Index: The key success index comprehensively considers the model's recall and false alarm rate, and is an important indicator for the overall evaluation of model performance. It represents the ratio of processes predicted as normal but actually being low-output processes out of all known low-output processes. Its calculation formula is:

[0098] Where: N TP +N FN +N FPThis includes all actual or predicted low-output events, i.e., all actual low-output events (including correctly identified and missed events) plus normal events that were incorrectly predicted as low-output events. TP indicates a successful prediction, meaning the model correctly predicted the occurrence of a low-output event; FP indicates a false alarm, meaning the model predicted a low-output period would occur, but it did not actually happen; FN indicates a missed event or prediction failure, meaning the event occurred, but the model did not predict it.

[0099] An ideal critical success index (CSI) value of 1 indicates that the model perfectly identifies all low-output processes without false alarms. A higher CSI suggests higher accuracy and reliability in predicting low-output processes.

[0100] (6) Fair skill score: The fair skill score reflects the proportion of both the forecast and the actual situation, and penalizes false reports and omissions.

[0101]

[0102]

[0103] In the formula: N is the total number of all possible cases, i.e., N TP +N FN +N FP +N TN N TP +N FP N represents the number of samples predicted as low-output processes. TP +N FN The number of samples that were actually low-output processes is denoted by ; the random hit number R is the number of samples predicted as low-output processes multiplied by the proportion of samples that were actually low-output processes, representing the number of samples that correctly predicted low-output processes under completely random guessing conditions.

[0104] ETS quantifies the improvement of an algorithm's prediction results compared to random guessing. The range is [-1, 1], where 0 indicates the algorithm's prediction performance is the same as random guessing (equally bad), and 1 indicates perfect prediction, meaning all positive and negative classes are correctly classified with no false positives or false negatives. Negative values ​​indicate worse performance than random guessing. ETS is "fair" because it considers not only the number of successful predictions but also adjusts for random successful predictions that may result from frequent events, thus providing a more equitable and reliable score.

[0105] 2.6 Case Study Verification: The data comes from 195 photovoltaic (PV) power stations in a northern province of my country, covering the provincial PV power output and daily guaranteed output prediction results from January 2021 to May 2022. This embodiment also uses gridded irradiance data provided by the CAMS Solar Radiation Service Database. Given the significant intermittent nature of PV power output, this embodiment selects data from 05:00 to 20:30 daily, with a 15-minute sampling interval for analysis. In the prediction of low-output processes, the input features include the temporal characteristics of the provincial PV daily guaranteed output and predicted power, and the gridded irradiance characteristics of the entire province. The spatial resolution of the irradiance data is 0.5°. The output is the prediction code (0 or 1) for the low-output process and its corresponding probability.

[0106] This embodiment employs the logistic regression algorithm and preliminarily observes its prediction performance on photovoltaic data from 195 photovoltaic power plants in a northern province, comparing it with the conventional "NWP-predicted power-low output calibration" method. To ensure fairness in comparing different prediction models, all models maintained consistent parameters except for the hyperparameters to be optimized. Parameter optimization for all algorithms involved selecting the optimal parameter combination through grid search and 5-fold cross-validation, with accuracy as the primary evaluation metric. The maximum number of training iterations was 500 for all algorithms, and all algorithms used identical training (first 80%) and test (last 20%) sets. The optimal parameter combination for the logistic regression model after grid search and 5-fold cross-validation was λ = 4.642, penalty = L1, and solvertype = liblinear. A low-output period identification model based on sample feature encoding was tested. The optimal hyperparameter combination was used for all five commonly used algorithms. A series of statistical indicators, including RR (Recall), AR (Accuracy), FAR (False Alarm Rate), MAR (Missing Alarm Rate), and CSI (Success Index), were used to quantitatively evaluate the results of different algorithms for identifying low-output photovoltaic periods under low-output thresholds at 5%, 10%, 15%, and 20%, respectively. Higher values ​​for RR, AR, and CSI indicate better identification performance, while lower values ​​for FAR and MAR indicate more accurate identification. CSI, a combination of RR and AR, reflects the overall probability of correct identification. Tables 1, 2, 3, and 4 show the comparison of the identification performance of different methods for intraday low-output periods under different low-output thresholds. Overall, the logistic regression model performed better across multiple thresholds.

[0107] Table 1 Comparison of recognition performance of low-output processes when the low-output threshold is 5%.

[0108]

[0109] Table 2 Comparison of recognition performance of low-output processes when the low-output threshold is 10%.

[0110]

[0111] Table 3 Comparison of recognition performance of low-output processes when the low-output threshold is 15%.

[0112]

[0113] Table 4 Comparison of recognition performance of low-output processes when the low-output threshold is 20%

[0114]

[0115] Compared to the conventional method "NWP-Predicted Power-Low Output Calibration," logistic regression improved recall (RR) and critical success index (CSI) to varying degrees across all thresholds, while reducing false negative rate (MAR). These results demonstrate that in identifying low-output photovoltaic periods, the "NWP-Low Output 01 Sequence" method, represented by logistic regression, exhibits better performance than the "NWP-Predicted Power-Low Output Sequence" method at all thresholds, providing strong support for improving identification accuracy and reliability.

[0116] This embodiment proposes a low-output period identification method based on sample spatiotemporal feature encoding. It clarifies the definition of low-output processes, constructs high-quality historical training samples, optimizes model parameters using grid search, and proposes a comprehensive evaluation index system. This method outperforms conventional methods in terms of accuracy, false alarm rate, and critical success index under different thresholds. It effectively improves recall and critical success index while reducing false alarm rate. Compared to conventional methods, it can more accurately identify low-output periods, providing a basis for dispatching to take proactive measures to address power supply and demand imbalances.

[0117] Research has found that the influencing factors on the intraday power supply guarantee capability of photovoltaic power also include the probability of photovoltaic power supply guarantee. Because the application value of related technologies in describing predictive uncertainty has not been fully explored, they cannot meet the needs of refined power grid dispatching decisions. Figure 11As shown, the prediction process for the probability of photovoltaic power supply guarantee in this embodiment specifically includes: SA, acquiring demand input data for the study area, wherein the demand input data includes current time information, regional gridded irradiance data for future time periods, and a lower limit for guaranteed photovoltaic power output for future time periods, wherein the lower limit for guaranteed photovoltaic power output for future time periods is a variable determined using the photovoltaic power prediction method described in Embodiment 1; SB, based on the demand input data, determining the discrete power prediction results for photovoltaic power at different confidence levels for future time periods using a trained quantile power prediction model, wherein the quantile power prediction model is a quantile regression based on a gradient-enhanced regression tree. The prediction model comprises: SC, which performs kernel density estimation on the discrete power prediction results to obtain the power prediction probability distribution density function for each future time; SD, which obtains a multi-timescale cumulative power distribution probability set based on the power prediction probability distribution density function for each future time, wherein the multi-timescale cumulative power distribution probability set includes the power prediction cumulative distribution function for each future time; and SE, which calculates the photovoltaic power supply guarantee probability set under a given power demand section based on the multi-timescale cumulative power distribution probability density set, wherein the photovoltaic power supply guarantee probability set includes the photovoltaic power supply guarantee probability under different time scales and different power demand in the future period. This embodiment can provide photovoltaic power supply guarantee probabilities for different times and different power / energy levels for grid dispatch, assisting in the formulation of scientific and reasonable power generation plans and improving the safety and stability of power system operation.

[0118] The following explains the specific process of the photovoltaic power supply guarantee probability extraction method.

[0119] Step 1: Use Gradient Boosting Regression Trees (GBRT) to perform quantile regression prediction and obtain discrete power prediction results at different confidence levels. Step 2: Calculate the power prediction probability distribution density function at each time point using kernel density estimation, and obtain the cumulative power distribution probability set across multiple time scales through integration. Step 3: For a given power demand profile (per unit value), calculate the survival function to obtain the photovoltaic power supply guarantee probability at different time scales and with different thresholds.

[0120] This embodiment studies a method for modeling the probability distribution of photovoltaic (PV) intraday power generation based on quantile regression and deep learning. It obtains discretized power probability prediction results at different confidence levels, and combines nonparametric distribution fitting interpolation and probability density function correction optimization. By introducing a PV guaranteed output curve for zero-probability elimination, the accuracy of the probability distribution is effectively improved. This method can provide grid dispatch with PV power supply guarantee probabilities at different times and with different power / energy levels, assisting in the formulation of scientific and reasonable power generation plans and improving the safety and stability of power system operation.

[0121] 3.1 Definition of Intraday Power Supply Guarantee Probability: Intraday power supply guarantee probability provides refined, comprehensive, and intuitive support decision-making information for power system operation and dispatch based on probability prediction, and quantifies power supply risk. In the actual dispatch decision-making process, grid operators are concerned with the probability that the photovoltaic power generation level at a certain moment can reach the power supply demand output level, or the risk that it cannot reach the power supply demand. To predict whether the power generation at a certain future moment can reach / meet a specific power supply demand, the complementary graph of the cumulative probability distribution, also known as the survival function S(x), can be used. S(x) represents the probability P(X>x) that the random variable X is greater than a certain specific value x. Mathematically, it can be expressed as: S(x)=P(X>x)=1-F(x) (19).

[0122] In the formula: F(x) is the CDF; S(x) can be obtained by subtracting the value of F(x) from 1. Based on the survival function at a specific time point, the probability that the actual power exceeds the power demand level for any given power supply can be found. Power system dispatching decisions only require inputting time t and power demand level x (per unit value) to determine the daily power supply guarantee probability by looking up a table.

[0123] Figure 12 The diagram illustrates the daily survival function of photovoltaic power generation across the province. Under the same daily power generation conditions, the survival probability of guaranteed daily output is lower than that of predicted and actual power generation, which is in line with expectations.

[0124] 3.2 Quantile Regression Prediction Based on Gradient Boosting Regression Tree: First, a photovoltaic probability distribution modeling method is studied based on quantile regression deep learning to obtain discrete power probability prediction results at different confidence levels. Gradient boosting regression tree is an ensemble learning method that improves prediction performance by integrating multiple weak learners (usually decision trees) to construct a strong learner. Combining gradient boosting regression tree and quantile regression, a quantile regression prediction model based on gradient boosting regression tree can be constructed.

[0125] Discrete power probability prediction results for renewable energy power plants are obtained based on multi-timescale stacking, such as... Figure 13 As shown, the probability density function at each time point is fitted using the kernel density method, and the cumulative distribution function is calculated by integration. The probability of power shortage at that time point is obtained by inputting the power demand level. Figure 14By aggregating the cumulative distribution functions at various time points, a multi-timescale cumulative distribution function set can be constructed. Using this function set, the probability of power shortage can be found at a given time and power demand level. This method can effectively assess and predict the power supply reliability of renewable energy power plants at different time scales, providing important basis for power system planning and scheduling. By calculating the complement graph to construct the multi-timescale survival function set, each point represents the probability that future power generation at that time can meet specific power demand.

[0126] 3.3 Zero-probability deletion: such as Figure 15 As shown, the zero-probability elimination method aims to improve the safety and reliability of the power system by adjusting the predicted probability density function to set the probability of output below the guaranteed output to zero. Specifically, based on the guaranteed output threshold, the portion of the predicted probability density function below this threshold is eliminated to ensure that the predicted output will not fall below the guaranteed output.

[0127] Let f(x) be the probability density of the original prediction at a certain moment, and P be the guaranteed output. min Zero-probability deletion is carried out through the following steps:

[0128] Adjusted density function The output will be lower than the guaranteed output P min The interval probability is set to 0:

[0129]

[0130] Adjusted density function Normalization is required to ensure the integral equals 1, resulting in a reasonable probability density function. The normalization process is as follows:

[0131] The integral of the adjusted unnormalized density function is calculated as follows:

[0132] The normalized density function is as follows:

[0133] Normalized adjusted probability density function This ensures that the probability of output below the guaranteed output is zero, while also guaranteeing that its integral is 1, thus becoming a valid probability density function.

[0134] 3.4 Evaluation index of power supply guarantee probability within 3 days: Based on the Brier score index, a power supply guarantee probability prediction and evaluation index system is constructed.

[0135] (1) Brier Score (BS): To accurately evaluate the performance of power supply guarantee probability prediction, the Brier score is used as the power supply guarantee probability prediction index. According to different actual needs, corresponding guarantee power sections are given, and the difference between the guarantee probability corresponding to the power supply demand section and the actual power supply situation is measured to evaluate the accuracy of the power supply guarantee probability prediction.

[0136]

[0137] In the formula: f t The probability that the actual output at time t meets the power supply demand; t The result at time t is the actual output meeting the power supply demand (1 for meeting, 0 for not meeting). The closer the Brier score is to 0, the better the power supply guarantee probability prediction performance; the closer it is to 1, the less accurate the power supply guarantee probability prediction.

[0138]

[0139] (2) Probability prediction accuracy (AR): Where: f t The probability that the actual output at time t meets the power supply demand; t The result of the actual power output at time t meeting the power supply demand (1 for meeting, 0 for not meeting). The closer AR is to 1, the better the power supply guarantee probability prediction performance; the closer it is to 0, the less accurate the power supply guarantee probability prediction.

[0140] 3.5 Numerical Example Verification

[0141] The data still uses 195 photovoltaic (PV) power stations in a northern province, covering the provincial PV output and intraday guaranteed output forecasts from January 2021 to May 2022, as well as gridded irradiance data provided by the CAMS solar radiation service database. In the intraday power supply guarantee probability prediction, the input features include the temporal characteristics of the provincial PV intraday guaranteed output and predicted power, and the gridded irradiance characteristics of the entire province. The model output is a set of power supply guarantee probabilities under different power demand conditions at different times of the day.

[0142] (1) Generating Confidence Intervals: In this embodiment, quantile regression prediction is performed on the power of the entire province to generate confidence intervals. For example... Figure 16 This is a graph showing the results of generating 95% confidence intervals from the test set data on January 19, 2021.

[0143] (2) Calculate the probability density of a given time point and power value: To further analyze the prediction results, assume that the dispatcher needs the probability that the photovoltaic output is greater than 0.8 (pu) at 12:30 on January 19, 2021. Figure 17The diagram shows the calculation of prediction results at different quantiles at 12:30, using kernel density estimation to fit the probability density function of the prediction results. By fitting the probability density function of the power prediction, the uncertainty of the predicted value can be visually observed. For example, the probability density is higher for predicted values ​​in the range of 0.8-0.9, indicating a higher confidence level for predictions within these ranges.

[0144] (3) Obtaining the photovoltaic guaranteed output curve and performing zero-probability elimination: The photovoltaic guaranteed output curve reflects the minimum output level of new energy sources under the premise of meeting the guarantee rate. It represents the minimum power value that the photovoltaic system guarantees energy output at a specific point in time. Therefore, by utilizing the constraint that the probability density of the portion of the power value below the guaranteed output is zero, the remaining portion is renormalized to make the total area equal to 1, thus adjusting the predicted probability density function. For example... Figure 18 The probability distribution function is the adjusted prediction result of different confidence levels at 12:30 on January 19, 2021. The photovoltaic guaranteed output at 12:30 is 0.996 (pu), and the probability density of output below 0.996 (pu) is set to 0.

[0145] like Figure 19 To obtain the CDF from the corrected zero-probability pruning PDF, with a power demand of 0.8 (pu), the photovoltaic guarantee probability S(0.8) = 1 - F(0.8) = 1, that is, the photovoltaic power supply guarantee probability at a given time level of 0.8 (pu) is 100%.

[0146] The above describes the method for calculating the guarantee probability at a single point. Generally, the guarantee probability prediction results for different power demand sections throughout the day are statistically analyzed. In the photovoltaic intraday power supply guarantee probability prediction section, a power demand section is set at an interval of 0.05 pu, constructing a multi-timescale multi-power demand section cumulative probability distribution function set and survival function set. The power supply guarantee probability prediction results for January 1, 2022 are as follows: Figure 20 As shown. (Through) Figure 20 The function set allows scheduling to select any time of day and power demand level, and then look up the corresponding power supply guarantee probability. For example... Figure 21 When the power demand level is selected as 0.2 pu Figure 17 The cross-section represents the probability of power supply guarantee at any time during the day. To evaluate the predicted power supply guarantee rate, the maximum power generation of the provincial photovoltaic cluster was reduced by 95% to 25% as the power demand cross-section for assessing the probability of power supply guarantee events for the provincial photovoltaic cluster. Table 5 analyzes the statistical results of the estimated Brier scores for the probability of power supply guarantee events for the provincial photovoltaic cluster during the day. The Brier scores of the photovoltaic cluster under different power demand cross-section conditions during the day are all below 0.03. Figure 22This is a graph showing the relationship between the Brier score and accuracy percentage of the test set and the power demand profile in Example 1 of this application; the prediction accuracy of the photovoltaic cluster under different power demand profile conditions during the day is higher than 88%, proving the accuracy of the proposed method in assessing the probability that the power generation will reach different power demand at different times in the future.

[0147] Table 5. Statistical Results of Brill's Score for Estimating the Daily Power Supply Guarantee Probability of Provincial Photovoltaic Clusters

[0148]

[0149] This embodiment studies a method for modeling the intraday probability distribution of photovoltaic power based on quantile regression deep learning. It obtains discretized power probability prediction results at different confidence levels, and obtains a continuous cumulative distribution function of the predicted power based on nonparametric distribution fitting and interpolation. The prediction results are then corrected and optimized using a probability density function. By introducing a photovoltaic guaranteed output curve, zero-probability pruning is applied to the probability distribution, effectively improving its accuracy.

[0150] Due to shortcomings in the adaptability, accuracy, uncertainty quantification, and low-output period identification of relevant forecasting technologies under different weather conditions, the accuracy and reliability of photovoltaic (PV) intraday power supply guarantee capabilities are affected, consequently impacting the accuracy and reliability of power dispatching plans. Therefore, addressing the issue that PV power forecasts sometimes exceed actual power, and the potential for overly aggressive dispatching plans based on these overestimated forecasts, this embodiment employs a power supply guarantee output forecasting method optimized by a multi-objective loss function. This method obtains a curve reflecting the minimum PV output with the goal of ensuring safety. Using deterministic forecast results as one of the input features, the multi-objective function comprehensively considers multiple penalty terms, including forecast accuracy, forecast positive deviation, smoothness, and response speed, to mine the actual power time-series characteristics, effectively extracting the PV power time-series patterns. Based on the guaranteed output lower limit result, the guaranteed power / electricity is obtained. To address situations where dispatching cannot proactively address power supply-demand imbalances during periods of low power output and weak power supply, this embodiment employs a method for identifying intraday low-output processes based on sample spatiotemporal feature coding. This method defines low-output processes, constructs historical samples of these processes, and considers the coupling relationship between low-output coding and meteorological characteristics. A pattern recognition model is then built, taking numerical weather prediction as input and low-output event coding and corresponding probabilities as outputs. Based on a method for calculating intraday power supply guarantee probability using quantile regression, the probability that photovoltaic power can meet the power supply needs under different grid power demand levels at a specific time is determined. By introducing a photovoltaic power guarantee output curve and performing zero-probability elimination on the probability distribution, the prediction accuracy is further improved.

[0151] Therefore, when formulating dispatch plans, power grid operators can optimize power generation plans based on more comprehensive, intuitive, and accurate forecast information, including guaranteed output curves, low-output period identification results, and intraday power supply guarantee probability. This provides a refined basis for power grid dispatching, allows for proactive responses to potential power supply and demand imbalances, and provides stronger support for power grid dispatching.

[0152] Example 2

[0153] This embodiment provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the photovoltaic intraday power supply guarantee capability prediction method described in Embodiment 1.

[0154] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0155] This embodiment uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application; at the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. In summary, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for predicting the daily power supply guarantee capacity of photovoltaic systems, characterized in that, The method for predicting the intraday power supply guarantee capacity of photovoltaic systems includes: The study predicts the minimum guaranteed output of photovoltaic (PV) power generation, the period of low PV output, and the probability of PV power supply guarantee for the future period in the research area. The minimum guaranteed output of PV power generation refers to the predicted minimum reliable power of PV power generation in the future period. The period of low output refers to the time when the first power difference is greater than the supplementary capacity of the PV backup unit. The first power difference refers to the difference between the PV power generation on a classic clear-sky day and the actual PV power generation. The probability of PV power supply guarantee refers to the probability that the predicted PV power generation in the future period can meet specific power supply needs. Based on the guaranteed output limit of photovoltaic power generation, the low output period of photovoltaic power generation, and the photovoltaic power supply guarantee probability, assess the photovoltaic intraday power supply guarantee capability in the future period; The process of predicting the guaranteed lower limit of photovoltaic power generation output specifically includes: Obtain multidimensional input parameters for the study area, including current time information, gridded irradiance data for future time periods, gridded wind speed data for future time periods, and historical actual photovoltaic power generation. Based on the multidimensional input parameters, the guaranteed output lower limit prediction model for future periods is used to determine the guaranteed output lower limit of photovoltaic power generation. The guaranteed output lower limit prediction model includes a convolutional neural network and a bidirectional long short-term memory network. The model is trained based on a multi-objective loss function, which is constructed based on a prediction accuracy penalty term, a prediction positive deviation penalty term, a smoothness penalty term, and a response speed penalty term. The prediction accuracy penalty term measures the prediction accuracy of the guaranteed output lower limit prediction model. The prediction positive deviation penalty term measures the magnitude of the prediction positive deviation, which is a variable determined based on the relationship between the guaranteed output lower limit of photovoltaic power generation and the actual photovoltaic power generation. The smoothness penalty term measures the smoothness of the guaranteed output lower limit of photovoltaic power generation over time. The response speed penalty term measures the sensitivity of the guaranteed output lower limit prediction model to changes in the actual photovoltaic power generation.

2. The method for predicting the daily power supply guarantee capacity of photovoltaic power generation according to claim 1, characterized in that, The prediction process for periods of low photovoltaic power generation specifically includes: Acquire target input data for the study area, wherein the target input data includes current time information, regional gridded irradiance data for future time periods, guaranteed output limit of photovoltaic power generation for future time periods, and photovoltaic power generation power on a classic clear-sky day; Based on the target input data, the low-output photovoltaic periods in the future are determined using a trained low-output period prediction model, wherein the low-output period prediction model is a machine learning model.

3. The method for predicting the daily power supply guarantee capacity of photovoltaic power generation according to claim 2, characterized in that, The training process of the prediction model for low-output periods specifically includes: A low-output training sample dataset is obtained, which includes historical time information, historical actual photovoltaic power generation, historical classic clear-sky day photovoltaic power generation, historical regional gridded irradiance data, low-output period coding sequence and non-low-output period coding sequence. The low-output period coding sequence is data obtained by coding historical low-output periods after merging processing. The non-low-output period coding sequence is data obtained by coding historical non-low-output periods. The historical low-output periods after merging processing are data obtained by merging adjacent low-output periods on the same historical day. Adjacent low-output periods refer to two low-output periods on the same historical day with a time interval less than or equal to a time interval threshold. The low-output period prediction model is trained using a grid search method based on the low-output training sample dataset.

4. The method for predicting the daily power supply guarantee capacity of photovoltaic power generation according to claim 2, characterized in that, The prediction model for low-output periods is a logistic regression model.

5. The method for predicting the daily power supply guarantee capacity of photovoltaic power generation according to claim 1, characterized in that, The process of predicting the probability of guaranteed photovoltaic power supply specifically includes: Obtain the demand input data for the study area, wherein the demand input data includes current time information, regional gridded irradiance data for future time periods, and the guaranteed output limit of photovoltaic power generation for future time periods; Based on the input data of the requirements, the discrete power prediction results of photovoltaics at different confidence levels in the future time period are determined by using the trained quantile power prediction model. The quantile power prediction model is a quantile regression prediction model based on gradient enhancement regression tree. Kernel density estimation is performed on the discrete power prediction results to obtain the power prediction probability distribution density function at each future time. A multi-timescale cumulative power distribution probability set is obtained based on the power prediction probability distribution density function at each future time point, wherein the multi-timescale cumulative power distribution probability set includes the power prediction cumulative distribution function at each future time point. Based on the multi-timescale cumulative power distribution probability density set, calculate the photovoltaic power supply guarantee probability set under a given power demand section, wherein the photovoltaic power supply guarantee probability set includes photovoltaic power supply guarantee probabilities under different time scales and different power demand in the future period.

6. The method for predicting the intraday power supply guarantee capacity of photovoltaic power generation according to claim 5, characterized in that, Based on the power prediction probability distribution density function at each future time point, a multi-timescale cumulative power distribution probability set is obtained, specifically including: The power prediction probability distribution density function for each future time period is filtered based on the guaranteed output limit of photovoltaic power generation in the future time period to obtain the power prediction probability distribution density function for each target future time period. The guaranteed output limit of photovoltaic power generation in the future time period is a variable determined using the photovoltaic power generation prediction method described in claim 1. Integrating the power prediction probability distribution density function of the target at each future time point yields a multi-timescale cumulative power distribution probability set.

7. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the photovoltaic intraday power supply guarantee capability prediction method according to claim 1.

Images