Photovoltaic power generation prediction method based on combination of regional features and weather features

Abstract

The invention discloses a photovoltaic power generation prediction method based on the combination of regional features and weather features, and relates to the technical field of photovoltaic power generation prediction, and the method comprises the following steps: constructing a regional feature hierarchical model which comprises a regional feature model, a seasonal feature model and a weather feature model, respectively quantifying the influence of regional parameters, seasonal changes and subdivided weather types on photovoltaic output; according to the method, the integrity, the accuracy and the consistency of the data are improved through hierarchical acquisition and preprocessing of the regional data, and a high-quality data basis is provided for model construction; according to the regional characteristic hierarchical modeling, influence factors are deeply analyzed from three dimensions of region, season and weather, and the accuracy and reliability of photovoltaic output prediction under different geographical conditions, season changes and weather conditions are remarkably improved; typical weather scene setting and simulation ensure that the model keeps high prediction precision under complex and changeable weather conditions.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation prediction technology, and in particular to a photovoltaic power generation prediction method based on a combination of regional characteristics and weather characteristics. Background Technology

[0002] With the increasing global demand for clean energy, photovoltaic (PV) power generation, as an important renewable energy source, has been widely adopted. However, PV power generation is affected by various factors such as location and weather, resulting in significant fluctuations and uncertainties in power output. This poses a significant challenge to the stable operation and dispatch of the power grid. Existing PV power generation forecasting methods often fail to fully consider the complexity of regional and weather characteristics, leading to low forecast accuracy and inability to meet practical application needs. Therefore, there is an urgent need for a method that can comprehensively consider regional and weather factors to improve the accuracy and reliability of PV power generation forecasting. Summary of the Invention

[0003] The purpose of this invention is to address the problem that existing photovoltaic power generation forecasting methods often fail to fully consider the complexity of regional and weather characteristics, resulting in low forecasting accuracy and inability to meet practical application needs. Therefore, this invention proposes a photovoltaic power generation forecasting method based on a combination of regional and weather characteristics.

[0004] To achieve the above objectives, the present invention employs a photovoltaic power generation prediction method based on a combination of regional and weather characteristics, comprising the following steps:

[0005] A regional characteristic hierarchical model is constructed, which includes a regional characteristic model, a seasonal characteristic model, and a weather characteristic model, respectively quantifying the impact of regional parameters, seasonal variations, and subdivided weather types on photovoltaic power output;

[0006] A typical weather scenario simulation mechanism is constructed based on the hierarchical model. Scenarios are generated through association rule mining and Monte Carlo simulation, and the scenario weights are evaluated by combining the analytic hierarchy process.

[0007] A dynamic optimization feedback mechanism is established, the model parameters are adjusted in real time using the Kalman filter algorithm, and the predictive adaptability is improved through transfer learning and online updates;

[0008] A hierarchical aggregation prediction model is constructed, which aggregates the prediction results of sub-regions according to the installed capacity weight to generate a prediction of the total power generation of the region.

[0009] Furthermore, the specific steps for constructing the regional feature model are as follows:

[0010] Using GPS and GIS fusion technology, the latitude and longitude of the sub-region are mapped into three-dimensional coordinates, and a terrain slope matrix is ​​constructed by combining DEM data;

[0011] The MODTRAN model is introduced to calculate the solar radiation correction coefficient, quantifying the effects of terrain shading and atmospheric refraction.

[0012] The gradient-enhanced random forest algorithm is used to quantify the influence weights of geographical factors such as altitude and slope on photovoltaic power output.

[0013] Furthermore, the specific steps for constructing the seasonal feature model are as follows:

[0014] The K-means clustering algorithm was used to divide meteorological data into seasons, and the meteorological characteristics and power generation patterns of each season were statistically analyzed.

[0015] An LSTM model with an attention mechanism is introduced to train each sub-region, and parameters such as seasonal average irradiance and temperature fluctuation coefficient are input to learn the dynamic effects of the season.

[0016] Furthermore, the specific steps for constructing the weather feature model are as follows:

[0017] We used CNN to extract features from satellite cloud images, combined with wavelet analysis of radar echo data, and used the RFE algorithm to filter weather features.

[0018] We construct a response model based on Transformer, design dedicated sub-modules for specific weather conditions such as rainstorms and typhoons, and introduce GNN to learn the spatial correlation of power plants.

[0019] Furthermore, the specific steps for constructing the typical weather scenario simulation mechanism are as follows:

[0020] The Apriori algorithm is used to mine the association rules of weather factors, and Monte Carlo scenes are generated by setting parameters such as precipitation intensity and duration.

[0021] A judgment matrix was constructed based on the analytic hierarchy process, and typical scenarios were selected using the degree of power generation impact and the probability of occurrence as indicators.

[0022] The model parameters were optimized through sensitivity analysis and adaptive gradient descent.

[0023] Furthermore, the dynamic optimization feedback mechanism comprises the following specific steps:

[0024] Real-time monitoring of error indicators such as MAE and RMSE, and dynamic adjustment of the weights of the hierarchical model through Kalman filtering;

[0025] Set a period to retrain the model and use transfer learning from similar regions to accelerate convergence;

[0026] Specific weather sub-models are updated online based on real-time weather and power generation data.

[0027] Furthermore, the specific steps for constructing the hierarchical aggregation prediction model are as follows:

[0028] Input the output parameters of the hierarchical model into the XGBoost model to calculate the aggregated power generation value of the sub-region;

[0029] Regional total forecast is generated by weighted summation based on installed capacity.

[0030] A heat map of the spatial distribution of power generation values ​​was visualized using GIS technology.

[0031] Furthermore, it also includes regional data stratified acquisition and preprocessing steps:

[0032] The target area is divided into sub-regions, and geographic information such as latitude, longitude and altitude is obtained through GIS to calculate the solar radiation correction coefficient.

[0033] Collect meteorological station data, satellite cloud images, power plant parameters, and historical power generation data;

[0034] By utilizing spatiotemporal interpolated gridded meteorological data, and by combining collaborative filtering with GAN to fill in missing power generation data, a unified spatiotemporal coordinate system is achieved.

[0035] Furthermore, the meteorological data includes hourly solar irradiance, temperature, humidity, wind speed, and detailed weather types, while the power station data includes installed capacity, component type, real-time power generation, and grid load data.

[0036] In summary, due to the adoption of the above-mentioned technology—a photovoltaic power generation prediction method based on a combination of regional and weather characteristics—the beneficial effects of this invention are:

[0037] This invention improves the integrity, accuracy, and consistency of data through hierarchical regional data collection and preprocessing, providing a high-quality data foundation for model construction. Hierarchical regional characteristic modeling deeply analyzes influencing factors from three dimensions: region, season, and weather, significantly improving the accuracy and reliability of photovoltaic power output prediction under different geographical conditions, seasonal variations, and weather conditions. Typical weather scenario settings and simulations ensure the model maintains high prediction accuracy under complex and variable weather conditions. Hierarchical aggregation prediction model construction enables accurate predictions from sub-regions to the entire region and provides visualization to assist decision-making. Dynamic model optimization and feedback ensure the model maintains high prediction accuracy and adaptability over the long term. The output and application of prediction results enhance the practicality and application value of the prediction results, helping to improve photovoltaic energy utilization efficiency, reduce operating costs, and enhance the stability and reliability of photovoltaic power plants in the power system. Attached Figure Description

[0038] Figure 1 A schematic diagram of the overall process of the present invention is shown. Detailed Implementation

[0039] The following will describe, with reference to the accompanying drawings of the embodiments of the present invention, a photovoltaic power generation prediction method based on a combination of regional and weather characteristics. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0040] The present invention will be further described below with reference to embodiments.

[0041] Example 1 (see) Figure 1 This invention provides a photovoltaic power generation prediction method based on a combination of regional and weather characteristics, comprising the following steps:

[0042] Step 1: Perform regional data stratification and preprocessing. Taking Hubei Province as an example, it is divided into multiple sub-regions, which can be based on prefecture-level cities or smaller grid units. Detailed geographic information for each sub-region is obtained using a Geographic Information System (GIS), including latitude and longitude, altitude, terrain slope, building density, etc. Based on GIS algorithms, the solar radiation correction coefficient for each sub-region is calculated to quantify the impact of terrain shading and atmospheric refraction on light intensity.

[0043] For multi-source data acquisition, real-time and historical data from multiple meteorological stations within the region are obtained through meteorological department APIs, including hourly solar irradiance, temperature, humidity, wind speed, and detailed weather types (sunny, cloudy, overcast, rainy, snowy, etc.). Satellite cloud images and radar echo data are also collected. Basic parameters (installed capacity, module type, inverter model), real-time power generation and historical power generation data of all photovoltaic power stations in Hubei Province are collected, as well as regional power grid node voltage and load distribution data.

[0044] In the data fusion and cleaning stage, spatiotemporal interpolation algorithms are used to grid the discrete meteorological station data to generate meteorological raster data covering the entire region; abnormal records in power generation data are removed, and a collaborative filtering algorithm based on neighboring stations combined with deep learning generative adversarial networks (GANs) is used to fill in missing data; the timestamps and spatial coordinate systems of geographic, meteorological, and power plant data are unified to construct a complete regional photovoltaic prediction dataset;

[0045] By dividing Hubei Province into multiple sub-regions and utilizing GIS to acquire geographic information and collect data from multiple sources, and employing spatiotemporal interpolation algorithms to grid meteorological data, remove anomalies and fill in missing data, and unify the spatiotemporal coordinate system, deep fusion and efficient processing of multi-dimensional data were achieved. This process effectively improved the completeness, accuracy, and consistency of the data, providing a high-quality data foundation for subsequent model construction, avoiding prediction biases caused by missing or anomalies, and enhancing the model's ability to represent actual power generation.

[0046] Step two, perform regional characteristic hierarchical modeling: Regional characteristic model: Based on the geographical parameters of sub-regions and the solar radiation correction coefficient, establish a regional solar radiation distribution model and calculate the theoretical solar irradiance of each sub-region. Use the random forest algorithm to analyze the correlation between geographical factors (elevation, slope) and historical power generation data, and quantify the influence weight of regional characteristics on photovoltaic output.

[0047] When constructing the regional feature model, the first step is to utilize the fusion technology of Global Positioning System (GPS) and Geographic Information System (GIS) based on the latitude and longitude coordinates of the sub-region. Map it to three-dimensional coordinates (X, Y, Z) in the Earth coordinate system using the three-dimensional space transformation formula:

[0048]

[0049] In the formula, (X,Y,Z) represents the latitude and longitude coordinates of the sub-region; (X,Y,Z) represents the three-dimensional coordinates of the sub-region in the Earth coordinate system after three-dimensional spatial transformation. Let be the radius of curvature of the circumpolar region, 'a' be the Earth's semi-major axis, 'e' be the Earth's first eccentricity, and 'h' be the altitude. Altitude data for each sub-region is obtained using a Digital Elevation Model (DEM). Then, using ArcGIS slope analysis tools, based on the slope calculation formula:

[0050]

[0051] In the formula, Slope is the slope of the sub-region; z is the elevation; and These are the partial derivatives of elevation z in the x and y directions, respectively, thus constructing a high-resolution terrain slope matrix.

[0052] When calculating the solar radiation correction factor, the atmospheric radiative transfer model MODTRAN is introduced, with its solar irradiance E... t The calculation formula is:

[0053] E t =E0×TOA×ATM×SC×AO

[0054] In the formula, E t 1 is solar irradiance; E0 is the solar constant; TOA is the solar radiation transmittance at the top of the atmosphere; ATM is the atmospheric absorption and scattering correction factor; SC is the solar zenith angle correction factor; AO is the aerosol optical thickness influence factor.

[0055] Using GIS shadow analysis tools, calculate the terrain occlusion correction factor O:

[0056]

[0057] In the formula, O is the terrain shading correction coefficient, where Shaded Area is the shaded area and Total Area is the total area;

[0058] The final solar radiation correction factor C is:

[0059] C = ATM × SC × AO × (1 - O)

[0060] In the formula, C is the final solar radiation correction factor, ATM is the atmospheric absorption and scattering correction factor, SC is the solar zenith angle correction factor, AO is the aerosol optical thickness influence factor, and O is the terrain shading correction factor.

[0061] When quantifying the weight of regional characteristics on photovoltaic power output, the Gradient Boosting Random Forest (GBRF) algorithm is used. Let the geographical factor feature vector be X = [x1, x2, ..., x...]. n (Including altitude, slope, etc.), the actual power generation value of historical power generation data is y, and the predicted value of the m-th decision tree is f. m If (X), then the prediction model of GBRF is:

[0062]

[0063] Where y represents the actual power generation value in historical data; This represents the predicted power generation value from the GBRF model; M is the number of decision trees; f m (X) is the predicted value of the m-th decision tree; γ m The weight of each tree;

[0064] By minimizing the loss function (such as mean square error) Where: N is the sample size, y i and The actual power generation value and the predicted power generation value of the i-th sample are respectively used to iteratively update the weight γ of each tree using the gradient descent method. m And a decision tree structure, thereby quantifying the weight of the impact of each geographical factor on photovoltaic output;

[0065] By utilizing GPS and GIS fusion technology, various calculation formulas, and the GBRF algorithm, the influence weight of regional characteristics on photovoltaic power output was accurately quantified, and a high-resolution terrain slope matrix and solar radiation distribution model were constructed. This model can fully consider the impact of geographical factors such as terrain shading and atmospheric refraction on light intensity, significantly improving the accuracy and reliability of photovoltaic power output prediction under different geographical conditions.

[0066] Seasonal Feature Model: Historical data is divided by season, and the meteorological characteristics (such as high temperature and humidity in summer and low irradiance in winter) and power generation patterns of different seasons in each sub-region are statistically analyzed. A season-specific Long Short-Term Memory (LSTM) network model is trained for each sub-region, inputting seasonal feature parameters (seasonal average irradiance, temperature fluctuation coefficient) and historical power generation data to learn the dynamic impact of seasonal changes on power generation values.

[0067] For seasonal division, the K-means clustering algorithm is used.

[0068] Let the meteorological data sample set be D = [d1, d2, ..., d]. N ], each sample (Including p meteorological indicators such as temperature and humidity), the cluster center vector is... (k = 1, 2, ..., K, where K is the number of clusters). Sample d i To the cluster center μ k The Euclidean distance is:

[0069]

[0070] In the formula: d(d i ,μ k ) is sample d i To the cluster center μ k Euclidean distance; d i Let μ be the i-th sample in the meteorological data sample set. k x is the cluster center vector; ij The j-th meteorological index in each sample; μ kj The cluster center vector μ k The j-th component in the data; p is the number of meteorological indicators contained in each meteorological data sample;

[0071] By iteratively updating the cluster centers, the sum of the distances from each sample to its respective cluster center is minimized:

[0072]

[0073] Among them, S k Let be the sample set in the k-th cluster, thus achieving seasonal division based on climate similarity.

[0074] An attention mechanism is introduced when training a season-specific LSTM model. Let the input sequence be X = [x1, x2, ..., x...]. T (X is an input sequence containing seasonal characteristic parameters and historical power generation data), and the attention weight a is obtained through attention calculation. t :

[0075]

[0076] Among them, a t For attention weights; e t =score(h t-1 ,x t The `score` function is used to calculate attention; it can take the form of a dot product, a weighted sum, etc. t-1 This represents the hidden state of the LSTM at the previous time step. The input X is then attention-weighted. a for:

[0077]

[0078] In the formula, X a It is the input after attention weighting;

[0079] X a Input the data into the LSTM model for training to minimize the predicted power generation value. The mean square error (MSE) between the actual power generation value (y) and the actual power generation value (y) is used to learn the dynamic impact of seasonal variations on the power generation value.

[0080] The K-means clustering algorithm was used for seasonal segmentation, combined with an LSTM model incorporating an attention mechanism. This effectively captured the meteorological characteristics and power generation patterns of different seasons in each sub-region, learning the dynamic impact of seasonal changes on power generation. This allows the model to adapt to power generation fluctuations caused by seasonal changes, achieving relatively accurate power generation forecasts in different seasons.

[0081] Weather Feature Model: This model performs refined classification of weather types, subdividing "rainy" into light rain, moderate rain, heavy rain, and torrential rain, and "partly cloudy" into less cloudy, partly cloudy, and overcast based on cloud cover. By combining satellite cloud imagery, radar echo intensity and texture, and lightning location data, dynamic features of each sub-weather type are extracted, such as cloud movement speed, precipitation particle size, and convective activity intensity. A Transformer-based weather-power generation response model is constructed, with dedicated sub-models or modules designed for each specific weather condition. For example, for torrential rain, an analysis module is added to analyze the relationship between precipitation intensity, duration, and the surface cleanliness of photovoltaic modules and power generation efficiency; for typhoon weather, the impact of wind speed and direction changes on the stability of photovoltaic supports, potential damage to power generation equipment, and power generation value is considered. The sub-models are trained using extensive historical weather data and corresponding power generation data, and an attention mechanism is introduced to ensure the model focuses on weather features that significantly impact power generation value.

[0082] In the refined classification of weather types, convolutional neural networks (CNNs) are used to extract features from satellite cloud images. Let the satellite cloud image be I∈R. H×W×C(H is the height, W is the width, and C is the number of channels). After passing through the convolutional layer Conv, the pooling layer Pool, and the fully connected layer FC, the feature vector f is obtained: f = FC(Pool(Conv(I))). Combining the radar echo data, wavelet analysis is used to decompose the radar echo signal s(t) into different frequency components:

[0083]

[0084] Where s(t) is the radar echo signal; J is the decomposition level; d j,k and c J,k For wavelet coefficients and scaling coefficients, ψ j,k (t) and Using wavelet functions and scaling functions, precipitation particle information is extracted by analyzing high-frequency and low-frequency features. Satellite cloud image features, radar echo features, and conventional meteorological data are fused, and features are then filtered using the recursive feature elimination method (RFE). Let the initial feature set be F, and the classifier be C. Features with low scores are continuously eliminated by calculating the feature importance score S: S = importance(C(F)) until a preset number of features is met, thus achieving refined weather type classification.

[0085] In constructing the Transformer-based weather-power generation response model, a strategy combining transfer learning and reinforcement learning is adopted. Taking the rainstorm weather sub-model as an example, in the pre-training stage, an initial model M0 is obtained by training with large-scale public data. In the fine-tuning stage, based on reinforcement learning, let the state space S represent the current weather and power generation data state, the action space A represent the model parameter adjustment actions, and the reward function R be the reciprocal of the prediction error.

[0086]

[0087] In the formula, s represents the state space, which refers to the current weather and power generation data state. It contains a series of feature information that can describe the current situation, such as the meteorological parameters at the current moment (temperature, humidity, solar irradiance, wind speed, weather type, etc.), the real-time power generation of the photovoltaic power station, and the equipment operating status. This information together constitutes the environmental state when the model makes decisions. The model decides what parameter adjustment actions to take based on the current state. a represents the action space, which is the model parameter adjustment action. In the Transformer-based weather-power generation response model, the action space a can be an operation that modifies various parameters in the model (such as the weights and biases of the neural network, the attention mechanism parameters in the Transformer architecture, etc.). By executing different parameter adjustment actions, the model tries to find a better combination of parameters to adapt to the constantly changing weather and power generation conditions. R(s,a) is the reward function. Mean Squared Error (MSE) is a commonly used indicator that measures the difference between predicted and actual power generation.

[0088] Actions are selected through a policy network π(a|s), and the network parameters are updated based on rewards from environmental feedback, enabling the model to adapt to heavy rain weather in the target area. Simultaneously, a graph neural network (GNN) is introduced, where the photovoltaic power station graph G = (V, E) has nodes v. i ∈V, and v i Indicates a power station; side (v) i ,v j )∈E, and (v i ,v j This represents the connection relationship between power stations. Node feature vectors. The update formula for layer l is:

[0089]

[0090] In the formula, Let W be the feature vector of node i at layer l+1; N(i) be the set of neighboring nodes of node i; and W be the feature vector of node i at layer l+1. l and b l Here, σ is a learnable parameter, and σ is the activation function. Let i be the feature vector of node j in the l-th layer; i and j are nodes.

[0091] By learning spatial relationships between power plants through GNNs, prediction accuracy can be improved;

[0092] This study utilizes CNN for feature extraction from satellite cloud images, wavelet analysis of radar echo data, and RFE (Resource-Free Evaluator) for feature filtering to achieve refined weather type classification. Combined with a Transformer-based weather-power generation response model, various learning strategies, and a Generative Neural Network (GNN), it provides an in-depth analysis of the impact of different weather conditions on photovoltaic power generation. Dedicated sub-models are designed for specific weather types, enabling the model to more accurately predict power generation under complex and variable weather conditions, thus improving the model's adaptability to weather factors and its prediction accuracy.

[0093] Step 3: Set up and simulate typical weather scenarios: Based on the characteristics of different regions and seasons, set up a variety of typical weather scenarios. In Hubei Province, set up a summer plum rain heavy precipitation scenario, setting different combinations of precipitation intensities (light rain, moderate rain, heavy rain, rainstorm), durations (1 hour, 3 hours, 6 hours, etc.), as well as humidity and light intensity;

[0094] The system simulates winter cold waves and strong winds, setting combinations of different wind speed levels (level 3, 6, 8, etc.), temperature drops (5℃, 10℃, etc.), and lighting conditions. A typhoon passage scenario is set up along the coast, setting combinations of weather parameters for different typhoon intensities (tropical depression, tropical storm, severe tropical storm, etc.), movement speeds, and paths. The constructed prediction model is used to simulate and predict each scenario, and the results are compared with actual historical data to evaluate the model's predictive performance under different scenarios. Model parameters and structure are then adjusted to optimize prediction accuracy.

[0095] In a typical weather scenario setting, the Apriori association rule mining algorithm is used. Let the weather factor itemset I = {i1, i2, ..., i...} m}, transaction set T = {t1, t2, ..., t n}, each transaction The support of itemset X, support(X), is:

[0096]

[0097] In the formula: I is the weather factor itemset; T is the transaction set; t j For each transaction; X is an itemset; support(X) is the support of itemset X;

[0098] By setting minimum support and minimum confidence thresholds, frequent itemsets and strongly associated rules are discovered. For example, for the rule... Its confidence level for:

[0099]

[0100] In the formula, for the rule It is its confidence level;

[0101] Based on the association rules mined, a scene is generated using Monte Carlo simulation. Taking a heavy rainfall scene during the summer plum rain season in Hubei Province as an example, the ranges of rainfall intensity P, duration D, humidity H, and light intensity L are set as [P...]. min ,P max ]、[D min D max ]、[H min H max ]、[L min ,L max i scene samples (P) are generated by random sampling. i D i H i ,L i ), i = 1, 2, ..., N.

[0102] The Analytic Hierarchy Process (AHP) is used to evaluate the scenario. A judgment matrix A is constructed, and its largest eigenvalue λ is calculated. max And the feature vector w, after passing the consistency test, yields the weight coefficients w1, w2, w3 for each evaluation index (such as the degree of impact on power generation, probability of occurrence, and similarity to historical data). Scenario S i Overall score (Score) i )for:

[0103] Score(S i = w1 × Impact(S) i )+w2×Probability(S i )+w3×Similarity(S i )

[0104] In the formula: Score(S i S represents the overall score for the scenario; i For the scene; Impact(S) i Probability (S) i Similarity(S) i ) are respectively scene S i The evaluation values ​​for the degree of impact on power generation, the probability of occurrence, and the similarity to historical data; w1, w2, and w3 are weighting coefficients;

[0105] Scenarios with high overall scores were selected as typical scenarios.

[0106] In scenario simulation prediction and optimization, scenario sensitivity analysis is introduced. Let the prediction model be f(x), where x = [x1, x2, ..., x]. n ] represents the scene parameter vector, for parameter x j Perform perturbation Δx j Calculate the sensitivity index S j :

[0107]

[0108] In the formula, x is the scene parameter vector; Δx j For parameter x j The perturbation; x′ is the parameter x j The perturbed vector; S j This is a sensitivity indicator.

[0109] Based on the sensitivity analysis results, an adaptive parameter adjustment strategy is adopted. Let the model parameter vector be θ. Based on the scene characteristics and sensitivity indicators, the parameters are updated using an optimization algorithm (such as stochastic gradient descent).

[0110]

[0111] In the formula, θ is the model parameter vector; θ k+1 θ k These are the parameter vectors after the (k+1)th and kth updates, respectively; η is the learning rate, and L(f(x; θ), y) is the loss function. By continuously simulating predictions, conducting sensitivity analysis, and adjusting parameters, the prediction accuracy of the model under different scenarios is optimized.

[0112] This study utilizes the Apriori algorithm to uncover association rules among weather factors, combines this with Monte Carlo simulation to generate scenarios, evaluates these scenarios using AHP (Adaptive High-Performance Hierarchy Process) and selects typical scenarios, and incorporates scenario sensitivity analysis and adaptive parameter adjustment strategies. This series of operations allows for the scientific setting of multiple typical weather scenarios based on historical data and actual weather conditions, comprehensively evaluating the model's predictive performance under different scenarios, and optimizing model parameters based on sensitivity analysis results. This ensures that the model maintains high predictive accuracy under complex and variable weather conditions, especially in extreme weather scenarios, providing strong decision support for photovoltaic power plants to cope with various weather conditions.

[0113] Step four involves constructing a hierarchical aggregation prediction model: The output parameters of the regional, seasonal, and weather characteristic models are input into the XGBoost prediction model of each sub-region to calculate the aggregated power generation value of the power plants within the sub-region. Based on the sub-region prediction results and installed capacity weights, a weighted sum is calculated to determine the total photovoltaic power generation of the entire region. GIS technology is used to visualize the spatial distribution of regional power generation values, generating a heat map to aid decision-making.

[0114] The method inputs the output parameters of regional, seasonal, and weather characteristic models into the XGBoost prediction model to calculate the power generation value of each sub-region. Then, based on the installed capacity weight, it calculates the total power generation of the entire region and uses GIS technology to visualize the spatial distribution of power generation values. This method achieves hierarchical aggregation prediction from sub-regions to the whole region, taking into account the characteristics of each sub-region while also providing a macroscopic understanding of the overall power generation situation of the region. It provides intuitive and comprehensive power generation data for grid dispatching and energy companies, assisting in the formulation of scientific and reasonable dispatching plans and investment plans.

[0115] Step 5: Dynamic model optimization and feedback. Establish a real-time prediction error monitoring mechanism, compare actual and predicted data, and calculate error indices (MAE, RMSE, etc.) for each sub-region. Utilize the Kalman filter algorithm to dynamically adjust the weight parameters of the regional, seasonal, and weather models based on the error. Set a model update cycle (e.g., weekly), and retrain the sub-models and aggregated models after accumulating new data. Introduce transfer learning techniques to transfer prediction experience from similar regions to the target region, accelerating model convergence. Simultaneously, feed real-time weather monitoring data and power generation data back into the model to update and optimize the prediction model online for specific weather conditions.

[0116] A real-time monitoring mechanism for prediction errors is established, the Kalman filter algorithm is used to dynamically adjust the model weight parameters, the model is retrained with a set update cycle, and transfer learning and online update optimization strategies are introduced. Through these measures, the model can track changes in power generation in real time, correct prediction errors in a timely manner, and fully utilize new data and experience from similar regions to continuously optimize its performance, ensuring that the model maintains high prediction accuracy and adaptability throughout long-term operation.

[0117] Step Six: Output and Application of Forecast Results: Output the overall predicted power generation for the region, the power generation distribution of each sub-region, and the power curve for the next 24 hours. Based on the forecast results, provide the grid dispatching department with a regional photovoltaic absorption capacity analysis report to assist in formulating peak-shaving plans; provide energy companies with regional photovoltaic investment planning suggestions to optimize power plant layout; and provide power plant operation and maintenance departments with equipment protection and operation and maintenance suggestions for specific weather conditions to reduce the adverse effects of weather on power plants.

[0118] It outputs detailed power generation forecasts and provides targeted analysis reports and recommendations to grid dispatching departments, energy companies, and power plant operation and maintenance departments. For grid dispatching, it offers absorption capacity analysis to assist in peak shaving; for energy companies, it provides investment planning suggestions to optimize layout; and for power plant operation and maintenance, it provides equipment protection and maintenance suggestions to reduce the impact of weather. This greatly enhances the practicality and application value of photovoltaic power generation forecasts, helps improve the utilization efficiency of photovoltaic energy, reduces operating costs, and enhances the stability and reliability of photovoltaic power plants in the power system.

[0119] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitution or modification made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the photovoltaic power generation prediction method and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A photovoltaic power generation prediction method based on a combination of regional and weather characteristics, characterized in that, Includes the following steps: A regional characteristic hierarchical model is constructed, which includes a regional characteristic model, a seasonal characteristic model, and a weather characteristic model, respectively quantifying the impact of regional parameters, seasonal variations, and subdivided weather types on photovoltaic power output; A typical weather scenario simulation mechanism is constructed based on the hierarchical model. Scenarios are generated through association rule mining and Monte Carlo simulation, and the scenario weights are evaluated by combining the analytic hierarchy process. A dynamic optimization feedback mechanism is established, the model parameters are adjusted in real time using the Kalman filter algorithm, and the predictive adaptability is improved through transfer learning and online updates; A hierarchical aggregation prediction model is constructed, which aggregates the prediction results of sub-regions according to the installed capacity weight to generate a prediction of the total power generation of the region.

2. The photovoltaic power generation prediction method based on a combination of regional characteristics and weather characteristics according to claim 1, characterized in that, The specific steps for constructing the regional feature model are as follows: Using GPS and GIS fusion technology, the latitude and longitude of the sub-region are mapped into three-dimensional coordinates, and a terrain slope matrix is ​​constructed by combining DEM data; The MODTRAN model is introduced to calculate the solar radiation correction coefficient, quantifying the effects of terrain shading and atmospheric refraction. The gradient-enhanced random forest algorithm is used to quantify the influence weights of geographical factors such as altitude and slope on photovoltaic power output.

3. The photovoltaic power generation prediction method based on a combination of regional characteristics and weather characteristics according to claim 1, characterized in that, The specific steps for constructing the seasonal feature model are as follows: The K-means clustering algorithm was used to divide meteorological data into seasons, and the meteorological characteristics and power generation patterns of each season were statistically analyzed. An LSTM model with an attention mechanism is introduced to train each sub-region, and parameters such as seasonal average irradiance and temperature fluctuation coefficient are input to learn the dynamic effects of the season.

4. The photovoltaic power generation prediction method based on a combination of regional characteristics and weather characteristics according to claim 1, characterized in that, The specific steps for constructing the weather feature model are as follows: We used CNN to extract features from satellite cloud images, combined with wavelet analysis of radar echo data, and used the RFE algorithm to filter weather features. We construct a response model based on Transformer, design dedicated sub-modules for specific weather conditions such as rainstorms and typhoons, and introduce GNN to learn the spatial correlation of power plants.

5. The photovoltaic power generation prediction method based on a combination of regional characteristics and weather characteristics according to claim 4, characterized in that, The specific steps for constructing a typical weather scenario simulation mechanism are as follows: The Apriori algorithm is used to mine the association rules of weather factors, and Monte Carlo scenes are generated by setting parameters such as precipitation intensity and duration. A judgment matrix was constructed based on the analytic hierarchy process, and typical scenarios were selected using the degree of power generation impact and the probability of occurrence as indicators. The model parameters were optimized through sensitivity analysis and adaptive gradient descent.

6. The photovoltaic power generation prediction method based on a combination of regional characteristics and weather characteristics according to claim 5, characterized in that, The dynamic optimization feedback mechanism comprises the following steps: Real-time monitoring of error indicators such as MAE and RMSE, and dynamic adjustment of the weights of the hierarchical model through Kalman filtering; Set a period to retrain the model and use transfer learning from similar regions to accelerate convergence; Specific weather sub-models are updated online based on real-time weather and power generation data.

7. The photovoltaic power generation prediction method based on a combination of regional characteristics and weather characteristics according to claim 6, characterized in that, The specific steps for constructing the hierarchical aggregation prediction model are as follows: Input the output parameters of the hierarchical model into the XGBoost model to calculate the aggregated power generation value of the sub-region; Regional total forecast is generated by weighted summation based on installed capacity. A heat map of the spatial distribution of power generation values ​​was visualized using GIS technology.

8. The photovoltaic power generation prediction method based on a combination of regional characteristics and weather characteristics according to claim 1, characterized in that, It also includes regional data stratified acquisition and preprocessing steps: The target area is divided into sub-regions, and geographic information such as latitude, longitude and altitude is obtained through GIS to calculate the solar radiation correction coefficient. Collect meteorological station data, satellite cloud images, power plant parameters, and historical power generation data; By utilizing spatiotemporal interpolated gridded meteorological data, and by combining collaborative filtering with GAN to fill in missing power generation data, a unified spatiotemporal coordinate system is achieved.

9. The photovoltaic power generation prediction method based on a combination of regional characteristics and weather characteristics according to claim 1, characterized in that, The meteorological data includes hourly solar irradiance, temperature, humidity, wind speed, and detailed weather types. The power station data includes installed capacity, component type, real-time power generation, and grid load data.

Images