A day-ahead prediction method for photovoltaic cluster power considering the identification of transition meteorological processes

Abstract

The present invention discloses a photovoltaic cluster power day-ahead prediction method that takes into account the identification of transitional meteorological processes, relating to the technical field of photovoltaic power prediction. The method comprises: utilizing an adaptive filtering method and a coded convolution scoring method to perform weather discrimination on the generated power, determining whether the current weather on the target day is abrupt or clear, constructing a corresponding weather forecast model based on the weather conditions, and constructing a photovoltaic cluster power day-ahead prediction model based on the two models, which has the dual functions of identifying meteorological processes and transitional points. The photovoltaic cluster power day-ahead prediction model is then used to predict the generated power of the photovoltaic cluster. By taking into account transitional and sudden weather events, the present invention can improve forecast accuracy and reliability.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power prediction, and in particular to a photovoltaic cluster power day-ahead prediction method taking into account the identification of transition meteorological processes. Background Art

[0002] China's new energy industry has been booming in recent years. As a clean energy source, photovoltaic power generation (PV) output is significantly affected by weather conditions, particularly cloud cover, resulting in intermittent and uncertain power. PV power forecasting, however, faces limitations in obtaining high-quality data, accurately describing meteorological factors, accurately tracking and predicting cloud trajectories, and optimizing models. These characteristics pose challenges to the stable operation of the power system and place higher demands on its regulatory capabilities. With the high proportion of renewable energy integrated into the power system, the stable operation of the power system faces significant challenges. High-precision power forecasting can help improve the absorption capacity of PV power generation, adjust power generation plans, and ensure the safe operation of the power system.

[0003] Although the pronounced cyclical nature of photovoltaic power output has led researchers to prioritize data on a daily basis to facilitate forecasting model construction and analysis, this approach fails to fully account for the nonlinear and irregular nature of weather changes. Weather fluctuations do not always follow a daily pattern but are influenced by a variety of factors, each with varying degrees of complexity across different regions. Consequently, current photovoltaic power forecasts lack accuracy and reliability. Summary of the Invention

[0004] The purpose of the present invention is to provide a photovoltaic cluster power day-ahead prediction method taking into account the identification of transition meteorological processes, so as to improve the prediction accuracy and reliability by taking the transition and mutation of weather into consideration.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A photovoltaic cluster power day-ahead prediction method taking into account the identification of transition meteorological processes includes:

[0007] Obtaining the generated power on the target day; the generated power includes a trend component and a fluctuation component;

[0008] Using an adaptive filtering method and a coded convolution scoring method to perform weather discrimination on the generated power, determining whether the current weather on the target day is sudden weather or sunny weather;

[0009] When the current weather on the target day is a sudden change, a turning point evaluation method is used to identify the sudden change point: if a sudden change point exists, a sudden change weather prediction model is constructed based on the sudden change characteristics of the earliest sudden change point; if no sudden change point exists, the current weather is regarded as sunny;

[0010] When the current weather on the target day is sunny, a non-mutational weather prediction model is constructed based on the static characteristics of the geographical location;

[0011] Based on the sudden weather prediction model and the non-sudden weather prediction model, a photovoltaic cluster power day-ahead prediction model with dual identification functions of meteorological processes and turning points is constructed, and the photovoltaic cluster power day-ahead prediction model is used to predict the power generation power of the photovoltaic cluster.

[0012] Optionally, the method of using an adaptive filtering method and a coded convolution scoring method to perform weather discrimination on the photovoltaic power to determine whether the current weather on the target day is sudden change weather or sunny weather specifically includes:

[0013] extracting the trend component and the fluctuation component of the generated power by using an adaptive filtering method;

[0014] Quantitatively score each component using the coded convolution scoring method to obtain a trend component score and a fluctuation component score;

[0015] Cluster analysis of each volatility score was performed using a boundary optimization method to determine the thresholds of key features of NWP data; the key features of NWP data included maximum daily rainfall, maximum daily cloud cover, and daily variation in cloud cover;

[0016] Weather forecasting is performed based on the key feature thresholds of the NWP data to determine whether the current weather is sudden change weather or sunny weather.

[0017] Optionally, the calculation process of the adaptive filtering method includes:

[0018] Based on the generated power, the window width is increased for periods with large fluctuation frequency but smooth fluctuation trend, and the window width is reduced for periods with turning fluctuation trend. The filtering formula is:

[0019]

[0020] Where x n+1 (i) is the i-th value of the (n+1)th filtered sequence, and the mean is calculated under the window of (id(i),i+d(i)); 2d(i)+1 is the filtering window corresponding to the i-th sequence value; x(j) is the j-th value of the filtered sequence;

[0021] Since the window length affects the degree of fluctuation of the filtered sequence, and the intensity of the sequence fluctuation is measured by difference, the adjustment length of the window is calculated using the following formula:

[0022] Δd n+1 (i) = int(kD(x n (i),2))

[0023] d n+1 (i) = Δd n (i)+d n (i)

[0024] Where D(·) is the second-order difference function; k is the adjustment coefficient; int() is the rounding function; Δd n (i) is the i-th value of the window correction value sequence of the n-th filtering; d n (i) is the i-th value of the window sequence after the n-th filtering.

[0025] Optionally, the method of using the boundary optimization method to perform cluster analysis on each fluctuation score to determine the key feature threshold of the NWP data includes the following specific steps:

[0026] The fluctuation scores were clustered using a boundary optimization method, with the upper quartiles of the daily maximum rainfall and the daily maximum cloud cover as corresponding thresholds, and the quartiles of the daily variation of cloud cover as corresponding thresholds.

[0027] Optionally, the calculation process of the turning point evaluation method includes:

[0028] Based on the turning score definition formula, the curvature of the fitted circle of the three data points is calculated to determine the degree of folding of the time series data points; wherein the turning score definition formula is:

[0029]

[0030] In the formula, (x center (i),y center (i)) is the coordinate of the center of the circle; xi is the i-th horizontal coordinate; yi is the i-th vertical coordinate.

[0031] Optionally, the calculation formula for the center coordinates is:

[0032]

[0033]

[0034] Where Δx k|k≠j is the difference between the horizontal coordinates of the other two points except j, Δy k|k≠j is the difference between the ordinates of the other two points except j; j is the jth horizontal coordinate; y j is the jth vertical coordinate.

[0035] Optionally, when the current weather on the target day is a sudden change, a turning point evaluation method is used to identify the sudden change point, and the specific process includes:

[0036] When the current weather on the target day is a sudden change, the ideal power generation period is determined based on the radiation duration, and each power generation period is evenly divided into three equal parts;

[0037] In each of the parts, a turning point evaluation method is used to identify mutation points. When a mutation point exists, the earliest mutation point is used as the mutation point of the photovoltaic cluster, and the trend correlation is calculated according to the time period divided by the turning point. The trend correlation is used as the mutation feature to construct a mutation weather prediction model.

[0038] Optionally, the sudden weather prediction model, the non-sudden weather prediction model and the photovoltaic cluster power day-ahead prediction model all adopt graph convolutional neural networks.

[0039] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0040] The present invention discloses a photovoltaic cluster power day-ahead prediction method that takes into account the identification of turning point meteorological processes. The method comprises using an adaptive filtering method and a coded convolution scoring method to perform weather discrimination on the power generation, determining whether the current weather on the target day is sudden change weather or sunny weather, constructing a corresponding weather forecast model according to the weather conditions, and constructing a photovoltaic cluster power day-ahead prediction model with dual identification functions of meteorological processes and turning points based on the above two models, and using the photovoltaic cluster power day-ahead prediction model to predict the power generation of the photovoltaic cluster.

[0041] This invention not only focuses on the transitions between weather types within a day, but also specifically on the degree of sudden change during severe weather conditions. Furthermore, the weather discrimination mechanism, which uses an inversion method to determine the thresholds for key features, enables the model to fully learn two different weather patterns, thereby avoiding accuracy loss in the event of sudden changes in weather type. Therefore, by taking into account weather transitions and sudden changes, the present invention can improve forecast accuracy and reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only 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.

[0043] Figure 1 This is a flow chart of the weather discrimination mechanism for determining key feature thresholds using the inversion method in this embodiment;

[0044] Figure 2 Construct a weather recognition and prediction comparison chart for the static chart in this embodiment;

[0045] Figure 3 Construct a weather recognition and prediction comparison chart for the dynamic chart in this embodiment;

[0046] Figure 4 This is a diagram showing power transition identification in this embodiment;

[0047] Figure 5 This is a comparison chart of the graph update prediction based on turning point identification in this embodiment; wherein, part (a) is sunny weather; part (b) is cloudy weather; part (c) is overcast weather; and part (d) is rainy weather.

[0048] Figure 6 This is a comparison chart of sunny day predictions in this embodiment;

[0049] Figure 7 This is a comparison chart of the overall method and common methods in this embodiment;

[0050] Figure 8 Schematic diagram of the logic flow of the photovoltaic cluster power day-ahead prediction method taking into account the identification of the transition meteorological process in this embodiment. DETAILED DESCRIPTION

[0051] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only 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 ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0052] The purpose of the present invention is to provide a photovoltaic cluster power day-ahead prediction method taking into account the identification of transition meteorological processes, so as to improve the prediction accuracy and reliability by taking the transition and mutation of weather into consideration.

[0053] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] like Figures 1-8 As shown, the present invention provides a photovoltaic cluster power day-ahead prediction method taking into account the identification of transition meteorological processes, comprising:

[0055] The generated power on the target day is obtained; the generated power includes a trend component and a fluctuation component.

[0056] The weather discrimination of the power generation power is performed using an adaptive filtering method and a coding convolution scoring method to determine whether the current weather on the target day is sudden change weather or sunny weather.

[0057] When the current weather on the target day is sudden change weather, the turning point is identified using the turning point evaluation method: if there is a sudden change point, a sudden change weather prediction model is constructed based on the sudden change characteristics of the earliest sudden change point; if there is no sudden change point, the current weather is regarded as sunny weather.

[0058] When the current weather on the target day is sunny, a non-mutational weather prediction model is constructed according to the static characteristics of the geographical location.

[0059] Based on the sudden weather prediction model and the non-sudden weather prediction model, a photovoltaic cluster power day-ahead prediction model with dual identification functions of meteorological processes and turning points is constructed, and the photovoltaic cluster power day-ahead prediction model is used to predict the power generation power of the photovoltaic cluster.

[0060] As a specific implementation of the above steps, the following steps are provided:

[0061] 1) Weather classification method considering adaptive filtering and coded convolution scoring

[0062] In order to adapt to the power mutation situation, an adaptive filtering method is proposed to extract the trend component and fluctuation component of each day. For the period with large fluctuation frequency but smooth fluctuation trend, the window width should be increased, while for the turning point of the fluctuation trend, the window width should be appropriately reduced to avoid the turning point offset. The filtering formula (1) is:

[0063]

[0064] Where x n+1 (i) is the i-th value of the (n+1)th filtered sequence, and the mean is calculated under the window of (id(i),i+d(i)); 2d(i)+1 is the filtering window corresponding to the i-th sequence value; x(j) is the j-th value of the filtered sequence;

[0065] The window length affects the degree of fluctuation of the filtered sequence. The intensity of the sequence fluctuation can be measured by difference, so the adjustment of the window length can be reasonably adjusted by difference. Δd is the adjustment length calculated by formula (2):

[0066] Δd n+1 (i) = int(kD(x n (i),2)) (2)

[0067] d n+1 (i) = Δd n (i)+d n (i) (3)

[0068] Where D(·) is the second-order difference function; k is the adjustment coefficient; int() is the rounding function; Δd n(i) is the i-th value of the window correction value sequence of the n-th filtering; d n (i) is the i-th value of the window sequence after the n-th filtering. Through multiple iterations, the window length is continuously adjusted to reduce the number of changes in trend code and the deviation value of peak and valley time points.

[0069] 2) Inversion method to determine the key NWP characteristic thresholds for clear and cloudy skies

[0070] Firstly, adaptive filtering is used to extract the trend and fluctuation components of daily power generation. Then, a convolutional fluctuation scoring method with an embedded coding mechanism is proposed to quantitatively analyze the volatility of power generation, thereby obtaining trend and fluctuation component scores. Finally, a boundary optimization method is applied to cluster analysis of three key features, thereby accurately classifying different weather types.

[0071] Three key features are extracted from NWP data: daily maximum rainfall, daily maximum cloud cover, and daily cloud cover variation. A quartile-based thresholding method is used to calculate feature thresholds for each weather classification. Specifically, the upper quartile of each key feature is used as the threshold. Given that cloud cover fluctuation is a key indicator for identifying anomalies in clear weather classification, the cloud cover variation threshold is set at the 2nd quartile. These feature thresholds are used to predict clear weather.

[0072] 3) Method for identifying mutation points under non-clear sky conditions

[0073] At the atmospheric level, NWP changes relatively slowly. Because NWP has a high sensitivity in identifying turning weather processes, a turning evaluation method for discrete sequence points is proposed to measure the degree to which the sequence deviates from the straight line in a local area. The turning score is defined as formula (4):

[0074]

[0075] The degree of folding of the time series data points is measured by calculating the curvature of the fitted circle of three data points. center (i),y center (i)) is the coordinate of the center of the circle, and the calculation formula is:

[0076]

[0077] Where Δx k|k≠j is the difference between the horizontal coordinates of the other two points except j, Δy k|k≠j is the difference between the ordinates of the other two points except j; j is the jth horizontal coordinate; y j is the jth vertical coordinate.

[0078] 4) Dynamic graph construction method based on mutation point switching

[0079] Select NWP search features according to steps 1) and 2) and use the key feature threshold standard to judge sudden weather changes.

[0080] For sudden weather changes, we analyze the duration of radiation to determine ideal power generation periods. These periods are then evenly divided into three equal parts. For each part, we search for the earliest sudden change point in the PV cluster as the cluster's sudden change point, as per step 3). We calculate trend correlations based on the time periods divided by the turning points, and use this correlation as a sudden change feature to construct a prediction model.

[0081] If no turning point can be identified, a forecast model will be constructed based on the static characteristics of the geographic location.

[0082] For non-mutated weather, a prediction model will be constructed based on the static characteristics of the geographical location.

[0083] 5) Simulation calculation

[0084] Simulation inputs: Analyze the measured electric field data to determine the total installed capacity of the field; the rated capacity of each photovoltaic field; and numerical weather forecasts for irradiance, rainfall, cloud cover, humidity, and air pressure. The data sampling interval is 15 minutes. Based on steps 1) to 4), obtain a real-time prediction of the photovoltaic power output for the entire field.

[0085] 6) Error analysis

[0086] Let P Mi is the actual average power in period i, P Pi is the predicted average power of period i, N is the total period of daily assessment, P cap is the wind farm startup capacity. Then, the normalized root mean square error is defined as formula (6):

[0087]

[0088] The coefficient of determination is defined as formula (7):

[0089]

[0090] According to step 5), input the simulation input quantity, and calculate the error between the predicted power calculated by the model and the measured power through the error evaluation standard formula (5), formula (6), and formula (7) in step 6) to obtain the prediction accuracy.

[0091] Based on the above technical solution, the following embodiments are provided.

[0092] In this embodiment, the method mainly includes the following steps: 1) extracting historical power fluctuation characteristics through adaptive filtering-based fluctuation characteristics, accurately classifying weather conditions, and analyzing the meteorological data differences between clear weather and sudden weather; 2) extracting key features using NWP data such as rainfall and cloud cover, determining key feature thresholds through inversion methods, and distinguishing clear weather from sudden weather; 3) under sudden weather conditions, selecting NWP search features, dividing the search intervals into multiple time periods, and searching for turning points in each interval; 4) constructing a sudden weather prediction model based on a dynamic switching graph for turning point identification; 5) constructing a clear weather prediction model based on a static geographic distance map. 6) Finally, establishing a photovoltaic power prediction model based on the dual identification of meteorological processes and turning points.

[0093] Specifically:

[0094] To maximize the accuracy of photovoltaic power forecasting, this paper primarily uses multiple prediction models to perform day-ahead photovoltaic power forecasts on a test set. The experimental analysis used power data from a photovoltaic cluster in Jilin Province, along with NWP irradiance, rainfall, cloud cover, humidity, and air pressure, with a temporal resolution of 15 minutes and a total installed capacity of 534.914 MW. The data spanned the period from 00:00 on January 2, 2019, to 23:45 on July 1, 2019. Data from January to May served as the training set, and data from June served as the test set.

[0095] Step 1: Extraction of fluctuation characteristics and weather classification method based on adaptive window filtering. Figure 1 As shown in the figure, adaptive window filtering is used to extract the trend and fluctuation components of daily power generation. The volatility of power generation is quantitatively analyzed using a convolutional fluctuation scoring method with an embedded coding mechanism, thereby obtaining trend and fluctuation component scores. A boundary optimization method is used to cluster the three key features, thereby accurately classifying different weather types.

[0096] Step 2: The inversion method is used to determine the threshold of key features to distinguish weather types. The NWP data of clear weather are analyzed, and three key features are extracted from it: maximum daily rainfall, maximum daily cloud cover, and daily change in cloud cover. The threshold setting method based on quartiles is selected to calculate the feature threshold from each type of data after weather classification. Specifically, the upper quartile of each key feature is used as the threshold. In view of the fact that cloud cover fluctuations are the key indicator for identifying anomalies in sunny day classification, the threshold of cloud cover change is set to the quartile. These feature thresholds are used to predict whether the weather is clear. When the value of any feature exceeds the preset threshold range, it is considered to be a sudden change in weather;

[0097] This example compares two models: one is an adjacency matrix constructed under static geographic conditions (Model 1), and the other is a daily adjacency matrix dynamically updated based on NWP regional divisions (Model 2). These two models verify the effectiveness of weather identification in forecasting.

[0098] Table 1 Forecast error evaluation index of weather identification effectiveness

[0099]

[0100] like Figure 2 As shown in Figure 3, it can be seen that in the case of severe weather and significant sudden changes in power generation, the prediction effect of the static model often cannot keep up with the changes in actual power. When performing power forecasting under the dynamic model, the overall prediction effect is much better than the static one. Since the daily update of the adjacency matrix may effectively identify the weather similarity between power plants on a daily basis, the improvement in the weather forecast effect is not significant. As shown in Table 1, the MAE accuracy of weather recognition prediction in both cases is reduced by an average of 0.3%. By comparing these methods, it is proved that the prediction model with weather type recognition is more excellent in prediction effect.

[0101] Step 3: Identify the turning point of sudden weather changes. Changes in the local meteorological environment are not always based on a daily cycle, but are affected by geographical location and show varying degrees of complexity. When considering meteorological conditions, the turning point of the weather must be taken into consideration. First, select NWP search features. The key NWP features are different in different weather conditions. For example, during rainfall, the absolute value and change value of rainfall are much more important for judging the location of the mutation point than other features, while on cloudy or overcast days, cloud cover dominates. Therefore, based on the key feature threshold standard, whichever key feature exceeds the threshold will be used as the mutation point judgment feature (rainfall is the highest priority, cloud cover is the second priority); then, divide the search interval into multiple time periods. This study determined the ideal power generation period by analyzing the radiation duration, and evenly divided these periods into four equal parts.

[0102] Finally, the turning point is identified in each search interval. The curvature of each time point of all stations is calculated based on the trend component of each key feature in each search period. The maximum curvature of each photovoltaic power station is taken as the mutation point, and the earliest mutation point in the photovoltaic cluster is the mutation point of the cluster. Figure 4 As shown in Figure 2, the mutation period of a certain day in the test set is divided into four periods (ad), which are relatively accurate.

[0103] Considering the situation where there are no turning points within the search interval, if no turning points are identified within a certain search interval, this indicates that no significant weather changes occurred during that period. For such periods, a prediction model is constructed based on the static characteristics of the geographical location. This approach ensures that periods with little change in power generation can be accurately identified and processed, thereby improving the accuracy and efficiency of the prediction.

[0104] Step 4: Build a sudden weather prediction model based on a dynamic switching graph for turning point identification. For forecasting weather with frequent turning points, a graph convolutional neural network (GCN) can be used to analyze fluctuations in NWP data to extract similarities between weather patterns. These similarity metrics are used to construct links between graph nodes and assign weights to these links. To accommodate intraday sudden weather changes, the regional divisions are updated daily, and moments of sudden change are identified. At these moments, the connectivity between regions is updated accordingly, and the local adjacency matrix and weights are adjusted.

[0105] Following step 3, select key features for identification. Use adaptive window filtering to extract fluctuations in the NWP key feature data for each PV field. Calculate trend correlations based on time periods divided by turning points. This correlation analysis serves as weight for constructing the connectivity matrix.

[0106] As shown in Table 2, under the most severe meteorological conditions, the forecast model uses three different construction methods: one is to construct a static adjacency matrix based on geographic location (method 1); the second is to construct a dynamic composition that is updated daily based on NWP data clustering (method 2); and the third is to construct a dynamic composition that is locally updated when a weather turning point is identified (method 3).

[0107] Table 2 Error evaluation table of prediction models constructed with different graphs

[0108]

[0109] like Figure 5 As shown in Figure 1, the prediction curve updated based on the NWP turning point identification map can effectively track sudden changes and turning points in the actual power curve. Because the adjacency matrix constructed statically based on geographic location cannot capture the dynamic changes in daily weather, the prediction results of Method 1 are poor. On the other hand, although the adjacency matrix construction method based on daily updates of NWP data clustering has certain adaptability in dealing with daily weather changes, its prediction model still shows certain limitations when dealing with sudden and sudden changes in weather. Therefore, this method of identifying turning points can more accurately capture the turning points of power generation, thereby optimizing the prediction model of the photovoltaic power generation system.

[0110] Step 5: Build a clear-weather forecast model based on the static geographic distance map. Smoothing the training power data with adaptive window filtering significantly improves forecast accuracy when used as training input. Clear-weather conditions, due to lower fluctuations, make it easier to filter out noise and interference to enhance model training, thereby ensuring accurate and reliable forecast results.

[0111] Table 3. Evaluation table of sunny day prediction error

[0112]

[0113] like Figure 6 As shown in Table 3, the prediction accuracy of the input data after preprocessing is very high. When training and predicting, the peak value of the unprocessed raw data deviates from the actual observed data and cannot accurately capture the peak value change of the real data.

[0114] Step 6: Establish a photovoltaic power prediction model based on dual identification of meteorological processes and turning points, and compare it with existing methods to verify the effectiveness of the overall method. The most commonly used prediction methods: Overall prediction (Method 1): The power is summed and predicted, and the output of training and testing is the overall power; Cumulative prediction (Method 2): Single station prediction, and then superimposed predicted power. Methods 1 and 2 are used for the most basic comparison to prove the effectiveness of existing methods. As well as prediction methods that consider weather scenarios and weather identification, Method 3: Cluster analysis is performed on daily NWP data to distinguish different weather types, and a special prediction model is constructed for each weather type. The model selects the most appropriate prediction model for weather identification and prediction by analyzing the cluster center of NWP data. Method 4: A weather division method based on historical power data is used to establish targeted prediction models for different weather conditions, realizing scenario-based prediction of photovoltaic power. Method 5: Clustering is performed based on the extracted power fluctuation characteristics to distinguish different weather types, and a special prediction model is established for each category. As Figure 7 As shown in the figure, the prediction results of different methods for several consecutive days are shown, and the prediction results under various weather conditions are shown in detail. The method proposed in this embodiment can perform stably and well under various weather conditions, and the prediction results are closely aligned with the trend of the actual data. The predicted power fluctuations of Method 3 and Method 5 are roughly consistent with the actual situation, but they fail to closely match the actual power. This shows that the model has preliminarily learned the characteristics under different weather conditions, but the learning and recognition of these characteristics are not deep enough, and the model shows instability in learning and recognizing the characteristics under different weather conditions.

[0115] Table 4 Error evaluation table of each prediction method

[0116]

[0117] As shown in Table 4, prediction methods that fail to consider similar day and weather identification perform poorly. This is because these methods fail to account for the unique characteristics of each PV field and the differences in characteristics under different weather conditions. In contrast, the method that uses a scenario-based prediction model significantly improves performance. Compared with the other five methods, the MAE and RMSE are reduced by an average of 4.96% and 7.15%, respectively, and the R² is increased by an average of 23.8%, demonstrating the effectiveness of our method.

[0118] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0119] The present invention is described in detail using specific examples. The above examples are only intended to help understand the core concept of the present invention. At the same time, those skilled in the art will appreciate that the specific implementation methods and scope of application may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A photovoltaic cluster power day-ahead prediction method taking into account the identification of transition meteorological processes, characterized by: include: Obtaining the generated power on the target day; the generated power includes a trend component and a fluctuation component; Using an adaptive filtering method and a coded convolution scoring method to perform weather discrimination on the generated power, determining whether the current weather on the target day is sudden weather or sunny weather; When the current weather on the target day is a sudden change, a turning point evaluation method is used to identify the sudden change point: if a sudden change point exists, a sudden change weather prediction model is constructed based on the sudden change characteristics of the earliest sudden change point; if no sudden change point exists, the current weather is regarded as sunny; When the current weather on the target day is sunny, a non-mutational weather prediction model is constructed based on the static characteristics of the geographical location; Based on the sudden weather prediction model and the non-sudden weather prediction model, a photovoltaic cluster power day-ahead prediction model with dual identification functions of meteorological processes and turning points is constructed, and the photovoltaic cluster power day-ahead prediction model is used to predict the power generation of the photovoltaic cluster; The adaptive filtering method and the coded convolution scoring method are used to perform weather discrimination on the generated power to determine whether the current weather on the target day is sudden change weather or sunny weather. The specific process includes: extracting the trend component and the fluctuation component of the generated power by using an adaptive filtering method; Quantitatively score each component using the coded convolution scoring method to obtain a trend component score and a fluctuation component score; Cluster analysis of each volatility score was performed using a boundary optimization method to determine the thresholds of key features of NWP data; the key features of NWP data included maximum daily rainfall, maximum daily cloud cover, and daily variation in cloud cover; Weather forecasting is performed based on the key feature thresholds of the NWP data to determine whether the current weather is sudden change weather or sunny weather.

2. The photovoltaic cluster power day-ahead prediction method taking into account the identification of transition meteorological processes according to claim 1 is characterized in that: The calculation process of the adaptive filtering method includes: Based on the generated power, the window width is increased for periods with large fluctuation frequency but smooth fluctuation trend, and the window width is reduced for periods with turning fluctuation trend. The filtering formula is: Where x n+1 (i) is the i-th value of the (n+1)th filtered sequence, and the mean is calculated under the window of (id(i),i+d(i)); 2d(i)+1 is the filtering window corresponding to the i-th sequence value; x(j) is the j-th value of the filtered sequence; Since the window length affects the degree of fluctuation of the filtered sequence, and the intensity of the sequence fluctuation is measured by difference, the adjustment length of the window is calculated using the following formula: Δd n+1 (i)=int(kD(x n (i),2)) d n+1 (i)=Δd n (i)+d n (i) Where D(·) is the second-order difference function; k is the adjustment coefficient; int() is the rounding function; Δd n (i) is the i-th value of the window correction value sequence of the n-th filtering; d n (i) is the i-th value of the window sequence after the n-th filtering.

3. The photovoltaic cluster power day-ahead prediction method taking into account the identification of transition meteorological processes according to claim 1 is characterized in that: The aforementioned method of using boundary optimization to cluster the fluctuation scores and determine the key feature thresholds of NWP data includes the following steps: The fluctuation scores were clustered using a boundary optimization method, with the upper quartiles of the daily maximum rainfall and the daily maximum cloud cover as corresponding thresholds, and the quartiles of the daily variation of cloud cover as corresponding thresholds.

4. The photovoltaic cluster power day-ahead prediction method taking into account the identification of transition meteorological processes according to claim 1 is characterized in that: The calculation process of the turning point evaluation method includes: Based on the turning score definition formula, the curvature of the fitted circle of the three data points is calculated to determine the degree of folding of the time series data points; wherein the turning score definition formula is: In the formula, (x center (i),y center (i)) is the coordinate of the center of the circle; x i is the i-th horizontal coordinate; y i is the i-th vertical coordinate.

5. The photovoltaic cluster power day-ahead prediction method taking into account the identification of transition meteorological processes according to claim 4 is characterized in that: The calculation formula of the center coordinates is: Where Δx k|k≠j is the difference between the horizontal coordinates of the other two points except j, Δy k|k≠j is the difference between the ordinates of the other two points except j; j is the jth horizontal coordinate; y j is the jth vertical coordinate.

6. The photovoltaic cluster power day-ahead prediction method taking into account the identification of transition meteorological processes according to claim 1 is characterized in that: When the current weather on the target day is a sudden change, a turning point evaluation method is used to identify the sudden change point. The specific process includes: When the current weather on the target day is a sudden change, the ideal power generation period is determined based on the radiation duration, and each power generation period is evenly divided into three equal parts; In each of the parts, a turning point evaluation method is used to identify mutation points. When a mutation point exists, the earliest mutation point is used as the mutation point of the photovoltaic cluster, and the trend correlation is calculated according to the time period divided by the turning point. The trend correlation is used as the mutation feature to construct a mutation weather prediction model.

7. The photovoltaic cluster power day-ahead prediction method taking into account the identification of transition meteorological processes according to claim 1 is characterized in that: The sudden weather prediction model, the non-sudden weather prediction model and the photovoltaic cluster power day-ahead prediction model all adopt graph convolutional neural networks.

Images