A photovoltaic power station prediction method and prediction model adaptive to weather changes

Abstract

This invention relates to the field of photovoltaic power prediction technology, specifically to an adaptive meteorological change-based photovoltaic power plant prediction method and model. The method acquires historical meteorological feature vectors to determine the static weather type for each day and calculates the change impact vector, which includes intraday specificity and abrupt changes between adjacent days. It calculates reference weights and performance weights for each static weather type to quantify data sparsity and inter-class uniqueness, thereby constructing the final reference weights for each day. These weighted clusters are then used to obtain conversion pattern categories, and a mapping relationship is established between weather conversion labels and conversion pattern categories. Finally, based on the meteorological forecast data for the day to be predicted, target labels are determined, and the prediction model is matched through the mapping relationship to output the photovoltaic power prediction results for the day to be predicted. This invention significantly improves the accuracy and robustness of photovoltaic power prediction under complex and extreme meteorological conditions through a two-level clustering system and a dual weight compensation mechanism.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power prediction technology, specifically to a photovoltaic power plant prediction method and prediction model that adapts to weather changes. Background Technology

[0002] With the accelerated global energy transition, photovoltaic (PV) power generation, due to its clean and renewable characteristics, has seen continuous and rapid growth in installed capacity, gradually becoming one of the core power sources in the new power system. However, PV power output exhibits significant intermittency, volatility, and randomness, and is greatly affected by meteorological factors (such as solar radiation, temperature, and cloud cover). For example, under complex meteorological conditions such as cloud cover and sudden weather changes, the output power of PV power plants may fluctuate drastically in a short period of time. This strong uncertainty poses severe challenges to power grid power balance scheduling, spinning reserve capacity allocation, and electricity market transactions. Therefore, achieving high-precision PV power plant power forecasting is of great significance for ensuring the safe and stable operation of the power grid and improving PV absorption capacity.

[0003] Currently, mainstream photovoltaic power prediction methods include physical modeling, statistical modeling, and artificial intelligence-based machine learning algorithms (such as Long Short-Term Memory Network (LSTM) and Support Vector Machine (SVM)). To improve prediction accuracy, similar day clustering is usually introduced as a key preprocessing technique. Its core logic is based on the principle that photovoltaic power output patterns are similar under similar weather conditions. Key features such as solar radiation, temperature, and humidity are selected, and the K-means clustering algorithm is used to divide historical data into several categories with similar weather characteristics (such as sunny days, cloudy days, and rainy days). Then, prediction models are built for different categories.

[0004] However, most existing similar-day clustering methods are based on static classification logic, that is, they only cluster based on the meteorological characteristics of the forecast day, ignoring the lag effect in the dynamic transition of weather types. This causes the model to fail to capture the nonlinear output deviation caused by weather transition, reducing the accuracy of photovoltaic power prediction. At the same time, insufficient attention is paid to low-density weather transition samples, which can easily lead to sparse but critical transition patterns (such as sudden rain and snow changes) being submerged or misclassified, resulting in poor generalization ability of the model when dealing with complex or extreme weather transition scenarios. In addition, the lack of refined quantification of the dynamic process of weather transition makes it difficult to accurately distinguish between smooth transitions and drastic changes, resulting in a lag or large deviation in the response of photovoltaic power prediction models at the moment of weather change, which makes it difficult to meet the needs of practical applications such as refined grid dispatching and electricity market trading. Summary of the Invention

[0005] To address the technical problem that existing similar-day clustering methods neglect weather transition lag effects and low-density sample characteristics, resulting in insufficient power prediction accuracy for photovoltaic power plants under complex meteorological conditions, the present invention aims to provide an adaptive photovoltaic power plant prediction method and model based on meteorological changes. The specific technical solution adopted is as follows: In a first aspect, one embodiment of the present invention provides a photovoltaic power plant forecasting method that adapts to weather changes, the method comprising the following steps: Obtain the daily meteorological feature vectors of the photovoltaic power station area within a historical time period; The static weather type of each day is determined based on meteorological feature vectors. The daily change influence vector is obtained based on the difference between the meteorological feature vector of each day and the meteorological feature vector of the day corresponding to its static weather type, as well as the difference between the meteorological feature vector of each day and the previous adjacent day. Based on the sparsity and discreteness of the meteorological feature vectors corresponding to each static weather type, the reference weight and performance weight of each static weather type are obtained; based on the reference weight and performance weight of the static weather type of each day and the previous adjacent day, the final reference weight of each day is obtained. Cluster the change impact vector using the final reference weight as the sample weight to obtain the conversion pattern category; combine the static weather type of each day with the previous adjacent day to construct the weather conversion label for each day, and establish the mapping relationship between the weather conversion label and the conversion pattern category; Based on the meteorological forecast data of the day to be predicted, the target weather transformation label is determined, and the prediction model is matched through the mapping relationship to output the photovoltaic power prediction result of the day to be predicted.

[0006] Furthermore, the method for obtaining the static weather type is as follows: The meteorological feature vectors are clustered using the K-means clustering algorithm to obtain clusters. Each cluster is then assigned a corresponding static weather type, thereby determining the static weather type for each day.

[0007] Furthermore, the method for obtaining the change influence vector is as follows: For any static weather type, the mean of all meteorological feature vectors corresponding to that static weather type is used as the category representation vector of that static weather type. For any day within a historical time period, the difference vector between the meteorological feature vector of that day and the category representation vector of its corresponding static weather type is taken as the first difference vector of that day. The difference between the meteorological feature vector of that day and the meteorological feature vector of the previous adjacent day is taken as the second difference vector of that day. The sum of the first difference vector and the second difference vector is used as the reference influence vector for that day. Based on all meteorological feature vectors within a historical time period, the global weight vector of the meteorological feature dimension is calculated using the entropy weight method. The vector obtained by element-wise multiplying the global weight vector and the reference influence vector is used as the change influence vector for that day.

[0008] Furthermore, the method for obtaining the reference weights is as follows: Obtain the number of meteorological feature vectors contained in each static weather type, and take the static weather type with the largest number as the main weather type; For any static weather type, the result of normalizing and positively shifting the difference in the number of meteorological feature vectors contained in the main weather type and the static weather type is used as the reference weight of the static weather type.

[0009] Furthermore, the method for obtaining the performance weights is as follows: For any static weather type, the distance between the static weather type and the central meteorological feature vector of the cluster corresponding to the main weather type is used as the inter-class reference distance for that static weather type. The average distance between all meteorological feature vectors and the central meteorological feature vector within the cluster corresponding to the static weather type is taken as the intra-cluster average distance of the static weather type. The negative correlation result of summing the intra-class average distances of the static weather type and the main weather type is used as the analysis weight of the static weather type. The normalized result of the product of the inter-class reference distance and the analysis weight for this static weather type is used as the performance weight for this static weather type.

[0010] Furthermore, the method for obtaining the final reference weight is as follows: For any day within a historical time period, the normalized result of the product of the reference weights of the static weather types of that day and the previous adjacent day is used as the reference degree of the transformation density of that day. The product of the performance weights of the static weather type of the day and the previous adjacent day is taken as the degree of weather transformation performance of the day. The product of the day's conversion density reference level and the conversion weather performance level is used as the final reference weight for that day.

[0011] Furthermore, the method for obtaining the mapping relationship is as follows: For any transformation mode category, the mean of the magnitudes of all change influence vectors in that transformation mode category is obtained and used as the reference influence index for that transformation mode category; Based on the magnitude of the reference impact index, the conversion mode categories are divided into different impact levels, and then a photovoltaic prediction model is constructed for the corresponding conversion mode category according to each impact level. Calculate the probability of each weather transition label falling into each transition mode category, remove mapping relationships with probabilities below a preset probability threshold, and establish a mapping relationship between weather transition labels and transition mode categories.

[0012] Furthermore, the method for matching the prediction model through the mapping relationship and outputting the photovoltaic power prediction result for the prediction date is as follows: When the target weather conversion label corresponds to only one conversion mode category, the photovoltaic prediction model corresponding to the current conversion mode category will be used as the target photovoltaic prediction model. Input the weather forecast data for the day to be predicted into the target photovoltaic prediction model, and output the photovoltaic power prediction result for the day to be predicted; When the target weather conversion label corresponds to at least two conversion mode categories, the photovoltaic prediction model corresponding to the conversion mode category at this time will be used as the reference photovoltaic prediction model. The output of each reference photovoltaic prediction model is obtained by inputting the meteorological forecast data for the day to be predicted into the model. This output serves as the reference predicted power for each reference photovoltaic prediction model. The result of negatively correlating and normalizing the historical errors of each reference photovoltaic prediction model is used as the fusion weight of each reference photovoltaic prediction model. The photovoltaic power prediction result for the date to be predicted is obtained by weighting and summing the reference predicted power using fusion weights.

[0013] Furthermore, the dimensions of the meteorological feature vector include total daily radiation, daily maximum temperature, daily radiation fluctuation coefficient, daily cloud cover rate of change, daily average humidity, daily precipitation, daily snowfall, and daily average wind speed.

[0014] Secondly, another embodiment of the present invention provides a photovoltaic power plant prediction model that adapts to weather changes. The system includes: a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of any of the above methods.

[0015] The present invention has the following beneficial effects: This invention first determines the daily static weather type based on meteorological feature vectors, which is beneficial for subsequent differentiated analysis of different basic weather patterns. Then, based on the difference in meteorological feature vectors between each day and the day corresponding to its static weather type, as well as the difference in meteorological feature vectors between each day and the previous adjacent day, it obtains the daily change impact vector, accurately reflecting the nonlinear hysteresis and dynamic mutation characteristics of weather state transitions on photovoltaic modules, preparing for subsequent identification and clustering of transformation patterns with similar dynamic impact characteristics. Furthermore, based on the sparsity and discreteness of the meteorological feature vectors corresponding to each static weather type, it obtains the reference weight and performance weight of each static weather type, accurately reflecting the data sparsity and inter-class separation (i.e., uniqueness) of each static weather type. To improve the sensitivity of the clustering algorithm to the identification of boundary samples and small sample transformation patterns, it further obtains the final reference weight for each day based on the reference weight and performance weight of the static weather type between each day and the previous adjacent day, accurately reflecting the daily change impact vector. The transformation process demonstrates the sparsity and specificity of the full data distribution. Furthermore, by clustering the change impact vector using the final reference weights as sample weights, transformation pattern categories are obtained. This facilitates the subsequent discovery of typical weather transformation dynamic patterns with physical consistency (such as radiation drop patterns and temperature hysteresis patterns). Next, the static weather types of each day are combined with the previous adjacent day to construct daily weather transformation labels, establishing a mapping relationship between weather transformation labels and transformation pattern categories. This accurately reflects the probabilistic correspondence between meteorological forecast labels and actual physical impact patterns, which is beneficial for subsequent adaptive model matching and fusion based on forecast data. Finally, the target weather transformation label is determined based on the meteorological forecast data for the forecast day. Through the mapping relationship, the prediction model is accurately matched, and the photovoltaic power prediction results for the forecast day are accurately output. This effectively improves the accuracy and robustness of photovoltaic power plant power prediction under complex and variable meteorological conditions, effectively meeting the needs of practical applications such as refined grid scheduling and electricity market trading. Attached Figure Description

[0016] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A schematic flowchart illustrating an adaptive weather change forecasting method for photovoltaic power plants, provided as an embodiment of the present invention; Figure 2 A structural diagram of a photovoltaic power plant prediction system for adaptive weather changes provided in one embodiment of the present invention; Figure 3 This is a schematic diagram of a computer device provided according to an embodiment of the present invention. Detailed Implementation

[0018] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of an adaptive weather-changing photovoltaic power plant prediction method and prediction model proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0020] The following description, in conjunction with the accompanying drawings, details the specific scheme of the adaptive meteorological change photovoltaic power plant prediction method and prediction model provided by this invention.

[0021] Example 1: This invention proposes an adaptive method for forecasting photovoltaic power plants based on weather changes. Please refer to [link / reference]. Figure 1 The diagram illustrates a schematic flowchart of an adaptive weather change forecasting method for photovoltaic power plants, provided by an embodiment of the present invention. The method includes the following steps: Step S1: Obtain the daily meteorological feature vectors of the photovoltaic power station area within the historical time period.

[0022] Specifically, to more accurately identify different weather types that affect photovoltaic (PV) power output, this embodiment first collects daily raw meteorological data of the PV power station area over a historical period through the sensor network of the PV power station or surrounding meteorological stations. This raw meteorological data includes radiation, temperature, cloud cover, humidity, and wind speed. Considering the varying degrees of impact of different meteorological factors on PV power station output, this embodiment extracts multi-dimensional feature indicators from different sensor data. Specifically, since radiation and temperature are core factors affecting PV power generation efficiency, the collected radiation and temperature data are used to calculate the daily total radiation and daily maximum temperature as the main features for distinguishing sunny weather. Considering that cloud thickness and its variations, as well as radiation fluctuations, directly affect the stability of PV power generation and are key criteria for distinguishing cloudy weather, the daily radiation fluctuation coefficient and daily cloud cover change rate are calculated as the main features for distinguishing cloudy weather. In addition, daily average humidity, daily precipitation, daily snowfall, and daily average wind speed are obtained as auxiliary features for distinguishing rain, snow, and strong winds. In this embodiment, the duration of the historical period is set to 3 years, and the end time of the historical period is 0:00 on the current day. The implementer can set the size of the historical period according to the actual situation, and there is no limitation here.

[0023] To improve data quality and facilitate more accurate subsequent analysis, this embodiment preprocesses the feature indicators within the historical time period. First, the Isolation Forest algorithm is used to detect anomalies in the feature indicators within the historical time period, removing the daily meteorological data corresponding to the feature indicators judged as anomalous (e.g., removing the top 1% of extreme data with the highest anomaly scores) to eliminate interference from sensor malfunctions or extreme emergencies on model training. Subsequently, the feature indicators after anomaly removal are subjected to Z-score standardization to eliminate the dimensional differences between different feature indicators, thereby generating standardized daily meteorological feature vectors. It should be noted that the dimensions of the meteorological feature vectors include total daily radiation, daily maximum temperature, daily radiation fluctuation coefficient, daily cloud cover change rate, daily average humidity, daily precipitation, daily snowfall, and daily average wind speed. The Isolation Forest algorithm and Z-score standardization are well-known techniques and will not be elaborated further.

[0024] Step S2: Determine the static weather type for each day based on the meteorological feature vector. Based on the difference between the meteorological feature vector of each day and the corresponding day of its static weather type, as well as the difference between the meteorological feature vector of the day and the previous adjacent day, obtain the daily change influence vector.

[0025] Specifically, in order to capture the fundamental patterns of photovoltaic power output under different meteorological conditions and construct a benchmark weather model, which is beneficial for subsequent differentiated analysis of different benchmark weather models, this embodiment first determines the static weather type of each day based on meteorological feature vectors. In reality, when different weather types switch, the previous weather type has a lag effect on the subsequent weather type, which leads to changes in the photovoltaic power output pattern. In order to quantify the dynamic inertia and abrupt change characteristics during the weather type switching process, so that the model can perceive and adapt to the power output fluctuations caused by weather abrupt changes, the daily change influence vector is obtained based on the difference between the meteorological feature vector of each day and the corresponding day of its static weather type, as well as the difference between the meteorological feature vector of the day and the previous adjacent day. This accurately reflects the nonlinear lag effect of weather state switching on photovoltaic modules, and prepares for subsequent identification and clustering of switching patterns with similar dynamic influence characteristics.

[0026] Preferably, in one feasible method of this embodiment, the static weather type is obtained as follows: First, the meteorological feature vectors are clustered using the K-means clustering algorithm to obtain clusters; wherein, one cluster corresponds to one static weather type. In this embodiment, the k value in the K-means clustering algorithm is set to 6, that is, 6 static weather types are set, namely sunny, cloudy, overcast, rainy, snowy, and extreme weather. Among them, sunny weather is characterized by the highest total daily radiation, extremely low cloud cover, and a smooth power generation curve that is close to the theoretical maximum value; cloudy weather is characterized by a relatively high but fluctuating total daily radiation and a moderate rate of change in cloud cover; overcast weather is characterized by a relatively low total daily radiation, high and continuous cloud cover, and a relatively low but stable overall power generation; rainy weather is characterized by significant daily precipitation, low radiation, and suppressed power generation efficiency; snowy weather is characterized by significant daily snowfall, which needs to be classified separately because snow accumulation may cause module shading (strong hysteresis effect) and ground reflection gain, and the photovoltaic response is completely different from that of rainy weather; extreme weather is characterized by... For days with extremely high radiation fluctuation coefficients or accompanied by strong winds (high daily average wind speed), representing drastic weather changes such as thunderstorms or severe convective weather, a corresponding static weather type is assigned to each cluster based on the performance of each of these static weather types. That is, after K-means clustering, the clusters are sorted and given fixed names based on the meteorological characteristic values ​​of the cluster centers (e.g., daily total radiation from highest to lowest) (e.g., type 1 for sunny days, type 6 for extreme weather), to ensure that the physical meaning of the static weather type remains consistent across different training batches; thus indirectly determining the static weather type for each day. The K-means clustering algorithm is a well-known technique and will not be elaborated further.

[0027] Preferably, in one feasible manner of this embodiment, the method for obtaining the change influence vector is as follows: Since different static weather type conversions have different effects on photovoltaic power output patterns, in order to analyze the subtle differences in meteorological conditions under different conversion modes, for any static weather type, the mean of all meteorological feature vectors corresponding to that static weather type is used as the category representation vector of that static weather type, thus representing the baseline meteorological feature state of that static weather type as a whole; for any day within a historical time period, the difference vector between the meteorological feature vector of that day and the category representation vector of its corresponding static weather type is used as the first difference vector of that day, accurately reflecting the intraday specific deviation of that day relative to the baseline state; Furthermore, considering the varying degrees of difference between different weather conditions, in order to capture the dynamic evolution characteristics of weather processes, the difference vector between the meteorological feature vector of the day and the meteorological feature vector of the previous adjacent day is obtained as the second difference vector of the day, which accurately reflects the degree of meteorological state change between the day and the previous adjacent day. In order to comprehensively characterize the combined influence of the static specificity and dynamic mutability of the day, so that subsequent clustering can take into account both static bias and dynamic trend, the sum of the first difference vector and the second difference vector is used as the reference influence vector of the day, which accurately reflects the potential influence of the day being affected by the lag of preceding weather and its own specificity. Considering the differences in information entropy of the impact of different meteorological feature dimensions on photovoltaic output (i.e., the greater the degree of variation, the more information the indicator contains), the global weight vector of each meteorological feature dimension is calculated using the entropy weight method based on all meteorological feature vectors within a historical time period. This accurately reflects the objective importance of each meteorological feature dimension in the impact assessment. It should be noted that the global weight vector and the reference impact vector have the same dimension. To accurately characterize the actual comprehensive impact intensity of each dimension feature after weighting on a given day, the vector obtained by element-wise multiplication of the global weight vector and the reference impact vector is used as the impact vector for that day. The entropy weight method is a well-known technique and will not be elaborated further.

[0028] This gives us the daily impact vector of the changes.

[0029] Step S3: Based on the sparsity and discreteness of the meteorological feature vectors corresponding to each static weather type, obtain the reference weight and performance weight of each static weather type; based on the reference weight and performance weight of the static weather type of each day and the previous adjacent day, obtain the final reference weight for each day.

[0030] Specifically, the distribution of historical meteorological data for photovoltaic power plants often exhibits significant density unevenness, meaning that the amount of data for common weather types such as sunny and cloudy days is much greater than that for low-frequency weather types such as snow and extreme weather. This data imbalance can further amplify the differences in sample density between different transition patterns when subsequently exploring weather transition patterns. For example, there are very few samples for transition patterns involving snow. If conventional clustering algorithms are used, it is easy for cluster centers to shift excessively towards high-density samples, thereby masking or ignoring low-density but crucial transition patterns for grid security. To overcome the limitations of uneven data distribution on the model's generalization ability, especially to improve the prediction accuracy for low-frequency key weather scenarios (such as snow and extreme weather), this embodiment introduces a clustering weighting mechanism. The core idea is based on the following two aspects: when the data density of a certain static weather type is lower (i.e., sparser), it is more likely to be ignored in conventional clustering, thus requiring a higher reference weight to compensate for this. On the other hand, the more discrete the feature distribution of a static weather type, the more likely that static weather type represents a unique physical process, that is, the more likely it is to form an independent transformation pattern, and thus requires assigning a higher performance weight to that static weather type. Therefore, this embodiment first obtains the reference weight and performance weight of each static weather type based on the sparsity and discreteness of the meteorological feature vectors corresponding to each static weather type, so as to accurately quantify the data sparsity and inter-class separation degree (i.e., uniqueness) of each static weather type.

[0031] Considering that weather transitions involve two adjacent days, and in order to accurately determine the comprehensive sample importance weight for each day, which is beneficial to strengthening the influence of sparsity and unique transition patterns in the subsequent clustering process, this embodiment obtains the final reference weight for each day based on the reference weight and performance weight of the static weather type of each day and the previous adjacent day. This accurately reflects the sparsity value and specific value of the transition process to which each day belongs in the full data distribution, and effectively improves the sensitivity of the subsequent clustering algorithm to the identification of boundary samples and small sample transition patterns.

[0032] Preferably, in one feasible implementation of this embodiment, the reference weight is obtained as follows: The number of meteorological feature vectors contained in each static weather type is obtained, and the static weather type with the largest number is taken as the primary weather type; then, for any static weather type, the result of normalizing and positively shifting the difference between the primary weather type and the number of meteorological feature vectors contained in that static weather type is used as the reference weight for that static weather type. The larger the reference weight, the fewer meteorological feature vectors the static weather type contains. To avoid low-frequency samples being overwhelmed by the features of high-frequency samples during clustering, the static weather type needs to have its weight increased to enhance its influence in the loss function. This embodiment... The difference between the number of meteorological feature vectors contained in the main weather type and the static weather type is normalized and positively shifted; where log is the logarithmic function with the natural constant as the base; x represents the difference between the number of meteorological feature vectors contained in the main weather type and the static weather type, and x is a non-negative number; The first preset constant is greater than 0, and is set in this embodiment. To be 0.1, to avoid Setting the reference weight to 0 ensures that all samples have positive reference weights, guaranteeing meaningful analysis later. The implementer can adjust this setting according to the specific circumstances. The size is not limited here.

[0033] At this point, the reference weights for each static weather type are obtained.

[0034] Preferably, in one feasible embodiment of this invention, the method for obtaining the performance weight is as follows: For any static weather type, the Euclidean distance between the central meteorological feature vector of the corresponding cluster of the static weather type and the main weather type is used as the inter-class reference distance of the static weather type. The larger the inter-class reference distance, the more significant the difference in feature distribution between the static weather type and the main weather type, and the higher its uniqueness. In order to more accurately analyze the compactness of the data distribution of the static weather type, the mean of the Euclidean distances between all meteorological feature vectors in the corresponding cluster of the static weather type and the central meteorological feature vector is used as the intra-class average distance of the static weather type. The smaller the intra-class average distance, the more concentrated the data distribution of the static weather type and the more stable the feature pattern. Furthermore, the result of negatively correlated summation of the intra-class average distances of the static weather type and the main weather type is used as the analysis weight of the static weather type. The larger the analysis weight, the smaller the sum of the intra-class dispersion of the static weather type and the main weather type, that is, the more compact the data of both types, which indirectly indicates that the inter-class reference distance is more meaningful. This embodiment is illustrated by... The sum of the intra-class average distances of the static weather type and the main weather type is negatively correlated, where norm is a linear normalization function and m is the sum of the intra-class average distances of the static weather type and the main weather type; additionally, it can also be done through... The sum of the intra-class average distances of static weather types and main weather types is negatively correlated; no specific restrictions are imposed here. This is a second preset constant, greater than 0, to avoid a denominator of 0. This embodiment sets... The value is 0.1, and the implementer can set it according to the actual situation. The magnitude of the distance is not limited here. Finally, the product of the inter-class reference distance and the analysis weight of the static weather type is linearly normalized, and the result is used as the performance weight of the static weather type. A larger performance weight indicates that the static weather type possesses both high inter-class separation and high intra-class compactness, and is a key weather model with significant independent characteristics. The methods for obtaining Euclidean distance and linear normalization are well-known techniques and will not be elaborated further.

[0035] At this point, the performance weights for each static weather type are obtained.

[0036] Preferably, in one feasible embodiment of this method, the final reference weight is obtained as follows: For any day within a historical time period, the product of the reference weights of the static weather types of that day and its preceding adjacent day is linearly normalized, and this product is used as the transformation density reference level for that day. The higher the transformation density reference level, the sparser the weather type data involved in the transformation process of that day is, and the more it needs to be prevented from being ignored by using high weights. The product of the performance weights of the static weather types of that day and its preceding adjacent day is used as the transformation weather performance level for that day. The higher the transformation weather performance level, the more unique and stable the weather type characteristics involved in the transformation process of that day are, and the greater the potential to form an independent transformation pattern. The value range of the transformation weather performance level is 0 to 1, because the value range of the performance weight is 0 to 1. In order to accurately characterize the comprehensive sample importance of that day, that is, the degree to which it needs to be given special attention in clustering, the product of the transformation density reference level and the transformation weather performance level of that day is used as the final reference weight for that day.

[0037] This completes the acquisition of the final reference weights for each day within the historical time period. It should be noted that the first day within the historical time period does not have a preceding adjacent day; therefore, the final reference weight for the first day within the historical time period is not acquired.

[0038] Step S4: Cluster the change impact vector using the final reference weight as the sample weight to obtain the conversion pattern category; combine the static weather type of each day with the previous adjacent day to construct the weather conversion label for each day, and establish the mapping relationship between the weather conversion label and the conversion pattern category.

[0039] Specifically, the known change impact vector accurately reflects the dynamic physical impact characteristics brought about by weather transitions. The final reference weights further illustrate the importance of each sample in identifying key transition patterns. Therefore, in this implementation, the change impact vectors are clustered using the final reference weights as sample weights to obtain transition pattern categories. This is beneficial for uncovering typical dynamic weather transition patterns with physical consistency (such as radiation drop patterns and temperature hysteresis patterns). To enable rapid matching of complex dynamic transition patterns based on simple weather labels during the prediction phase, the static weather types of each day and its preceding adjacent day are combined to construct a weather transition label for each day. For example, the weather transition label for day i is... ,in, The static weather type for day i-1. Let i be the static weather type for day i. Then, establish a mapping relationship between weather transition labels and transition mode categories to accurately reflect the probabilistic correspondence between weather forecast labels (such as sunny to cloudy) and actual physical influence patterns (such as radiation fluctuation intensity), thus preparing for subsequent adaptive model matching and fusion prediction based on weather forecast data.

[0040] Preferably, in one feasible embodiment of this invention, the method for obtaining the transformation pattern category is as follows: First, a weighted K-means clustering model is constructed, using the daily change impact vector as input samples and the corresponding final reference weights for each day as sample weights input into the model. During the clustering iteration process, a centroid update strategy is used to adjust the sample weights; that is, samples with larger weights have a greater attraction for the position of the cluster centroid, thereby ensuring that the clustering results can preferentially capture sparse but critical transformation patterns. This embodiment uses the elbow method to determine the optimal number of clusters, and finally labels the output clusters as different transformation pattern categories. Each transformation pattern category represents a group of weather transformation processes with similar dynamic physical impact characteristics (such as similar radiation change trends and similar temperature hysteresis). The weighted K-means clustering algorithm and the elbow method are well-known techniques and will not be described in detail here.

[0041] Preferably, in one feasible embodiment, the mapping relationship is obtained as follows: for any conversion mode category, the mean of the magnitudes of all change influence vectors in that conversion mode category is obtained as the reference influence index of that conversion mode category; the larger the reference influence index, the stronger the dynamic impact of weather conversion on photovoltaic output under that conversion mode (such as severe fluctuations in radiation); in order to adapt prediction models of different complexities to dynamic impacts of different intensities (e.g., using more complex deep learning models for high-impact levels), the conversion mode category is further divided into different impact levels based on the magnitude of the reference influence index, specifically: when the linearly normalized reference influence index is... When the impact index is greater than a preset first threshold, the corresponding conversion mode category is classified as a high-impact level; when the linearly normalized reference impact index is greater than or equal to a preset second threshold and less than or equal to a preset first threshold, the corresponding conversion mode category is classified as a medium-impact level; when the linearly normalized reference impact index is less than a preset second threshold, the corresponding conversion mode category is classified as a low-impact level. In this embodiment, the preset first threshold is set to 0.7 and the preset second threshold is set to 0.4, so that the classification results can evenly cover the three typical dynamic impact scenarios of strong, medium and weak. Implementers can set the preset first threshold and the preset second threshold according to the actual situation, which is not limited here. Then, a photovoltaic prediction model is constructed according to the corresponding conversion mode category for each influence level. Specifically, historical meteorological data (i.e., meteorological feature vectors) belonging to the conversion mode category and the corresponding photovoltaic power data are used as the training set. For high influence levels, a deep neural network model based on LSTM or Transformer is constructed to capture complex nonlinear features. For medium influence levels, a model based on Support Vector Machine (SVM) or Random Forest is constructed. For low influence levels, a lightweight model based on linear regression or shallow network is constructed to optimize computational efficiency while ensuring prediction accuracy. In order to determine how to select the appropriate conversion mode based on weather forecast labels, the probability of each weather conversion label falling into each conversion mode category is statistically analyzed. Specifically, the proportion of samples clustered into each conversion mode category in all samples of a certain weather conversion label (e.g., sunny to cloudy) is calculated. Considering that some low-probability mapping relationships may be caused by data noise or abnormal samples (e.g., occasional misclassification), mapping relationships with probabilities lower than the preset probability threshold are removed to ensure the establishment of an accurate mapping relationship between weather conversion labels and conversion mode categories. This embodiment sets a preset probability threshold of 0.05 to make the mapping relationship more robust and avoid low-probability noise interfering with the model matching decision. Implementers can set the size of the preset probability threshold according to the actual situation, and it is not limited here. Among them, deep neural network models such as LSTM or Transformer, models such as support vector machine (SVM) or random forest, and lightweight models of linear regression or shallow networks are all well-known technologies and will not be described in detail.

[0042] Step S5: Determine the target weather transition label based on the meteorological forecast data of the day to be predicted, match the prediction model through the mapping relationship, and output the photovoltaic power prediction result of the day to be predicted.

[0043] Specifically, to achieve accurate prediction of future photovoltaic power output, this embodiment first obtains weather forecast data (including features such as total daily radiation and daily maximum temperature) for the predicted day and weather data (or weather forecast data) for the adjacent day before the predicted day through a weather forecast interface. Based on the weather forecast data for the predicted day, the static weather type of the predicted day is determined using the K-means clustering model in step S2 (i.e., using pre-trained cluster centers); similarly, the static weather type of the adjacent day before the predicted day is determined. Subsequently, the static weather type of the adjacent day before the predicted day is combined with the static weather type of the predicted day (e.g., sunny-partly cloudy) to construct the target weather conversion label for the predicted day.

[0044] Next, based on the established mapping relationship between weather transition labels and transition mode categories, the corresponding transition mode category for the target weather transition label is found. In practical applications, due to the complexity of weather systems, a target weather transition label may correspond to one or more transition mode categories (for example, sunny to cloudy may mostly belong to the medium impact level, but there are also a few cases that belong to the high impact level). Therefore, this embodiment, based on the target weather transition label of the day to be predicted, matches the prediction model through the mapping relationship and outputs the photovoltaic power prediction result for the day to be predicted. This is beneficial for achieving adaptive and accurate matching between the model and the dynamic characteristics of weather, maximizing prediction performance.

[0045] Preferably, in one feasible way of this embodiment, the method for outputting the photovoltaic power prediction result of the day to be predicted is as follows: when the target weather conversion label corresponds to only one conversion mode category, it indicates that the dynamic characteristics of the weather conversion process are relatively simple and stable. At this time, the photovoltaic prediction model corresponding to the unique conversion mode category is directly used as the target photovoltaic prediction model. Input the meteorological forecast data (i.e., meteorological feature vector) of the day to be predicted into the target photovoltaic prediction model, and the output result is the photovoltaic power prediction result of the day to be predicted; When a target weather transition label corresponds to at least two transition mode categories, it indicates that the weather transition process has multiple possible dynamic evolution paths (i.e., the transition state is unstable or there are boundary effects). To improve the robustness of the prediction and avoid misjudgment by a single model, all photovoltaic prediction models associated with the corresponding transition mode categories are used as reference photovoltaic prediction models. The output of each reference photovoltaic prediction model is obtained by inputting the meteorological forecast data (i.e., meteorological feature vector) of the forecast date into each model, and this output serves as the reference predicted power for each model. To integrate the advantages of each model, the historical errors of each reference photovoltaic prediction model are negatively correlated and normalized, and this result is used as the fusion weight for each model. This embodiment uses... The historical errors mentioned above are negatively correlated and normalized, where w represents the historical errors and norm is a linear normalization function. Finally, the reference predicted power is weighted and summed using the fusion weights, and the result is the photovoltaic power prediction result for the day to be predicted.

[0046] It should be noted that if the target weather conversion label does not have a corresponding conversion mode in the mapping relationship, the general photovoltaic prediction model corresponding to the static weather type of the day to be predicted will be used as the target model for prediction.

[0047] Through the aforementioned adaptive matching and fusion strategy, this invention can not only handle conventional and stable weather scenarios, but also effectively cope with complex and ever-changing weather transition scenarios, significantly improving the accuracy and stability of photovoltaic power plant power prediction.

[0048] In summary, this embodiment obtains historical meteorological feature vectors to determine the static weather type for each day and calculates the change impact vector, which includes intraday specificity and abrupt changes between adjacent days. It calculates the reference weight and performance weight for each static weather type to quantify data sparsity and inter-class uniqueness, thereby constructing the final reference weight for each day. This weighted clustering of the change impact vector yields the transformation pattern category, and a mapping relationship is established between weather transformation labels and transformation pattern categories. Finally, based on the meteorological forecast data for the day to be predicted, the target label is determined, and the prediction model is matched through the mapping relationship to output the photovoltaic power prediction result for the day to be predicted. This invention significantly improves the accuracy and robustness of photovoltaic power prediction under complex and extreme meteorological conditions through a two-level clustering system and a dual weight compensation mechanism.

[0049] Example 2: This invention also proposes an adaptive weather-changing photovoltaic power plant forecasting system; please refer to [link / reference]. Figure 2The diagram illustrates a structure of an adaptive weather-changing photovoltaic power plant prediction system according to an embodiment of the present invention. The system includes: a data acquisition module 10, a change influence vector acquisition module 20, a final reference weight acquisition module 30, a mapping relationship acquisition module 40, and a photovoltaic power prediction module 50.

[0050] The data acquisition module 10 is used to acquire the daily meteorological feature vectors of the photovoltaic power station area within a historical time period.

[0051] The change impact vector acquisition module 20 is used to determine the static weather type of each day based on the meteorological feature vector, and to obtain the change impact vector of each day based on the difference between the meteorological feature vector of each day and the meteorological feature vector of the day corresponding to its static weather type, as well as the difference between the meteorological feature vector of the day and the previous adjacent day.

[0052] The final reference weight acquisition module 30 is used to obtain the reference weight and performance weight of each static weather type based on the sparsity and discreteness of the meteorological feature vector corresponding to each static weather type; and to obtain the final reference weight of each day based on the reference weight and performance weight of the static weather type of each day and the previous adjacent day.

[0053] The mapping relationship acquisition module 40 is used to cluster the change impact vector with the final reference weight as the sample weight to obtain the conversion mode category; it combines the static weather type of each day with the previous adjacent day to construct the weather conversion label for each day, and establishes the mapping relationship between the weather conversion label and the conversion mode category.

[0054] The photovoltaic power prediction module 50 is used to determine the target weather transformation label based on the meteorological forecast data of the day to be predicted, match the prediction model through the mapping relationship, and output the photovoltaic power prediction result of the day to be predicted.

[0055] It should be noted that the system provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the computer equipment can be divided into different functional modules to complete all or part of the functions described above. In addition, the photovoltaic power station prediction system and the photovoltaic power station prediction method embodiment that are adapted to weather changes provided in the above embodiments belong to the same concept, and their specific implementation process can be found in the method embodiment, which will not be repeated here.

[0056] Example 3: This invention also proposes an adaptive weather-changing photovoltaic power plant prediction model. This model includes a memory and a processor. The memory stores executable program code, and the processor calls and executes this executable program code to perform the adaptive weather-changing photovoltaic power plant prediction method provided in the embodiments of this application. Specifically, the model can be a chip, component, or module. The chip may include a connected processor and memory; the memory stores instructions, and when the processor calls and executes the instructions, the chip can perform the adaptive weather-changing photovoltaic power plant prediction method provided in the above embodiments.

[0057] In addition, this embodiment also protects a computer device; please refer to [link to relevant documentation]. Figure 3 The computer device includes a memory 401, a processor 402, and a computer program 403 stored in the memory 401 and running on the processor 402. When the processor 402 executes the computer program 403, the computer device can execute any of the aforementioned adaptive weather change photovoltaic power plant prediction methods.

[0058] Example 4: The present invention also provides a computer-readable storage medium storing computer program code, which, when run on a computer, causes the computer to execute the aforementioned related method steps to implement the adaptive weather change photovoltaic power plant prediction method provided in the above embodiments.

[0059] Example 5: The present invention also provides a computer program product that, when run on a computer, causes the computer to perform the aforementioned related steps to realize the adaptive weather change photovoltaic power plant prediction method provided in the above embodiments.

[0060] In this embodiment, the device, computer-readable storage medium, computer program product, or chip are all used to execute the corresponding methods provided above. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods provided above, and will not be repeated here.

[0061] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0062] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A photovoltaic power plant forecasting method that adapts to meteorological changes, characterized in that, The method includes the following steps: Obtain the daily meteorological feature vectors of the photovoltaic power station area within a historical time period; The static weather type of each day is determined based on meteorological feature vectors. The daily change influence vector is obtained based on the difference between the meteorological feature vector of each day and the meteorological feature vector of the day corresponding to its static weather type, as well as the difference between the meteorological feature vector of each day and the previous adjacent day. Based on the sparsity and discreteness of the meteorological feature vectors corresponding to each static weather type, the reference weight and performance weight of each static weather type are obtained; based on the reference weight and performance weight of the static weather type of each day and the previous adjacent day, the final reference weight of each day is obtained. Cluster the change impact vector using the final reference weight as the sample weight to obtain the conversion pattern category; combine the static weather type of each day with the previous adjacent day to construct the weather conversion label for each day, and establish the mapping relationship between the weather conversion label and the conversion pattern category; Based on the meteorological forecast data of the day to be predicted, the target weather transformation label is determined, and the prediction model is matched through the mapping relationship to output the photovoltaic power prediction result of the day to be predicted. The method for obtaining the reference weights is as follows: Obtain the number of meteorological feature vectors contained in each static weather type, and take the static weather type with the largest number as the main weather type; For any static weather type, the result of normalizing and positively shifting the difference in the number of meteorological feature vectors contained in the main weather type and the static weather type is used as the reference weight of the static weather type. The method for obtaining the performance weights is as follows: For any static weather type, the distance between the static weather type and the central meteorological feature vector of the cluster corresponding to the main weather type is used as the inter-class reference distance for that static weather type. The average distance between all meteorological feature vectors and the central meteorological feature vector within the cluster corresponding to the static weather type is taken as the intra-cluster average distance of the static weather type. The negative correlation result of summing the intra-class average distances of the static weather type and the main weather type is used as the analysis weight of the static weather type. The normalized product of the inter-class reference distance and the analysis weight of the static weather type is used as the performance weight of the static weather type. The method for obtaining the final reference weight is as follows: For any day within a historical time period, the normalized result of the product of the reference weights of the static weather types of that day and the previous adjacent day is used as the reference degree of the transformation density of that day. The product of the performance weights of the static weather type of the day and the previous adjacent day is taken as the degree of weather transformation performance of the day. The product of the day's conversion density reference level and the conversion weather performance level is used as the final reference weight for that day.

2. The photovoltaic power plant forecasting method for adaptive weather changes as described in claim 1, characterized in that, The method for obtaining the static weather type is as follows: The meteorological feature vectors are clustered using the K-means clustering algorithm to obtain clusters. Each cluster is then assigned a corresponding static weather type, thereby determining the static weather type for each day.

3. The photovoltaic power plant forecasting method for adaptive weather changes as described in claim 1, characterized in that, The method for obtaining the change impact vector is as follows: For any static weather type, the mean of all meteorological feature vectors corresponding to that static weather type is used as the category representation vector of that static weather type. For any day within a historical time period, the difference vector between the meteorological feature vector of that day and the category representation vector of its corresponding static weather type is taken as the first difference vector of that day. The difference between the meteorological feature vector of that day and the meteorological feature vector of the previous adjacent day is taken as the second difference vector of that day. The sum of the first difference vector and the second difference vector is used as the reference influence vector for that day. Based on all meteorological feature vectors within a historical time period, the global weight vector of the meteorological feature dimension is calculated using the entropy weight method. The vector obtained by element-wise multiplying the global weight vector and the reference influence vector is used as the change influence vector for that day.

4. The photovoltaic power plant forecasting method for adaptive weather changes as described in claim 1, characterized in that, The method for obtaining the mapping relationship is as follows: For any transformation mode category, the mean of the magnitudes of all change influence vectors in that transformation mode category is obtained and used as the reference influence index for that transformation mode category; Based on the magnitude of the reference impact index, the conversion mode categories are divided into different impact levels, and then a photovoltaic prediction model is constructed for the corresponding conversion mode category according to each impact level. Calculate the probability of each weather transition label falling into each transition mode category, remove mapping relationships with probabilities below a preset probability threshold, and establish a mapping relationship between weather transition labels and transition mode categories.

5. The photovoltaic power plant forecasting method for adaptive weather changes as described in claim 4, characterized in that, The method for matching the prediction model through the mapping relationship and outputting the photovoltaic power prediction result for the date to be predicted is as follows: When the target weather conversion label corresponds to only one conversion mode category, the photovoltaic prediction model corresponding to the current conversion mode category will be used as the target photovoltaic prediction model. Input the weather forecast data for the day to be predicted into the target photovoltaic prediction model, and output the photovoltaic power prediction result for the day to be predicted; When the target weather conversion label corresponds to at least two conversion mode categories, the photovoltaic prediction model corresponding to the conversion mode category at this time will be used as the reference photovoltaic prediction model. The output of each reference photovoltaic prediction model is obtained by inputting the meteorological forecast data for the day to be predicted into the model. This output serves as the reference predicted power for each reference photovoltaic prediction model. The result of negatively correlating and normalizing the historical errors of each reference photovoltaic prediction model is used as the fusion weight of each reference photovoltaic prediction model. The photovoltaic power prediction result for the date to be predicted is obtained by weighting and summing the reference predicted power using fusion weights.

6. The photovoltaic power plant forecasting method for adaptive weather changes as described in claim 1, characterized in that, The dimensions of the meteorological feature vector include total daily radiation, daily maximum temperature, daily radiation fluctuation coefficient, daily cloud cover rate of change, daily average humidity, daily precipitation, daily snowfall, and daily average wind speed.

7. A photovoltaic power plant prediction model adapted to weather changes, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the photovoltaic power plant prediction method for adaptive weather changes as described in any one of claims 1-6.

Images