Deep learning irradiance numerical mode prediction correction method and system introducing clustering

Abstract

The invention provides a deep learning irradiance numerical mode prediction correction method and system introducing clustering, and the method comprises the steps: enabling information sequences under the same generalized weather type to be classified into the same cluster through introducing a clustering algorithm; the variational mode decomposition algorithm can decompose historical data time sequence signals into various intrinsic mode functions, redundant information is removed, and effective features are obtained; the feature learning task is optimized, so that deep learning model training and error correction can be more effectively carried out subsequently.

Description

Technical Field

[0001] This invention relates to the field of irradiance correction technology, specifically to a deep learning-based method and system for correcting numerical irradiance forecasts by incorporating clustering. Background Technology

[0002] Currently, the world needs to vigorously develop non-fossil energy, accelerate the development of wind and solar power, and comprehensively promote the large-scale development and high-quality development of wind and solar power. In recent years, new energy power generation, represented by wind and solar power, has experienced explosive growth. As of 2023, taking China as an example, the cumulative installed capacity of solar power generation has reached approximately 610 GW, making it the second largest power source in the country. However, with the increase in installed capacity of solar power generation, the power output of solar power generation is affected by factors such as irradiance, resulting in fluctuations and random changes, which will lead to a further decline in solar energy utilization.

[0003] To mitigate the adverse effects of photovoltaic (PV) power generation instability on the power grid and the waste of solar energy resources, PV power generation forecasting technology is an effective solution. Accurate forecasting of PV output facilitates advance planning of power dispatch, allowing for the rational allocation of power plant output, thereby ensuring grid security, improving power quality, and guaranteeing the stable and reliable operation of the power system. Furthermore, China is actively promoting electricity marketization and has already implemented electricity spot trading in some regions. With the further development of the electricity market mechanism, ultra-short-term PV power generation forecasting technology will become even more important, promoting the regulation of electricity market pricing strategies and maximizing economic benefits.

[0004] Currently, using NWP numerical weather prediction data as input for prediction correction models is a widely adopted method. Correction processing of NWP solar radiation can effectively improve photovoltaic power prediction results. Compared to traditional methods such as data assimilation and model output statistics, many researchers are increasingly applying machine learning and deep learning methods for prediction bias correction. Deep learning models possess powerful representation learning capabilities, learning higher-level abstract features from large-scale data, which helps extract potential information and patterns from meteorological data, thus improving numerical model bias correction. To compensate for the shortcomings of single correction methods, many researchers combine the advantages of different correction methods, such as combining statistical methods in the data preprocessing stage with deep learning correction methods, to improve the accuracy of irradiance correction. While mining the hidden spatiotemporal features in model forecast data, strengthening the spatiotemporal correlation of training data enhances the overall bias correction capability of the network structure, thereby improving the accuracy of the corrected forecast. However, the correction results are highly sensitive to changes in weather conditions. Most existing NWP GHI correction models do not consider separate modeling for different weather conditions, yet NWP GHI errors can vary significantly depending on weather conditions. Meanwhile, weather factors and irradiance contain discrete variables, and there is a nonlinear mapping relationship between them. Current correction methods are insufficient in dealing with the random and discrete relationships between variables. Summary of the Invention

[0005] To address the shortcomings of existing technologies in handling random and discrete relationships between variables, this invention proposes a deep learning-based method for correcting numerical irradiance model forecasts by incorporating clustering, comprising: Obtain the station observation radiation data, station observation meteorological elements, numerical weather prediction model forecast radiation data and numerical weather prediction model forecast meteorological elements for a set period before the correction time, as well as the numerical weather prediction model forecast radiation data and numerical weather prediction model forecast meteorological elements on the correction date. Clustering algorithms were used to cluster the observed radiation data and meteorological elements from the stations, and to classify them into multiple weather types. The field observation radiation data were processed using a variational mode decomposition algorithm to obtain multiple intrinsic mode functions; Based on the weather type, the station observation meteorological elements, numerical weather prediction model forecast radiation data, numerical weather prediction model forecast meteorological elements and various intrinsic mode functions, as well as the numerical weather prediction model forecast radiation data and numerical weather prediction model forecast meteorological elements on the correction date, are input into the pre-trained irradiance numerical model forecast correction model to obtain the numerical weather prediction model forecast radiation correction result on the correction date. The irradiance numerical model forecast correction model is a CNN-LSTM network with an encoder-decoder structure. The irradiance numerical model forecast correction model is trained using historical field observation irradiance data, field observation meteorological data, numerical weather prediction model forecast irradiance data, and numerical weather prediction model forecast meteorological elements.

[0006] Preferably, the training process of the irradiance numerical model forecast correction model includes: Acquire historical data on field observations of radiation, meteorological elements observed at the field, radiation data and meteorological elements predicted by numerical weather prediction models, and divide them into training and validation sets according to a set ratio. Clustering algorithms were used to cluster the observed radiation data and meteorological elements from the stations, and to classify them into multiple weather types. The field observation radiation data were processed using a variational mode decomposition algorithm to obtain multiple intrinsic mode functions; For each weather type, a CNN-LSTM network with an encoder-decoder structure is trained using the following inputs: meteorological elements observed at the station for a set duration before the correction time in the training set, radiometric data predicted by the numerical weather prediction model, meteorological elements predicted by the numerical weather prediction model, various intrinsic mode functions, and radiometric data predicted by the numerical weather prediction model at the correction time. The output is the radiometric data observed at the station on the correction date. The model is then validated using a validation set to obtain the irradiance numerical model forecast correction model.

[0007] Preferably, the CNN-LSTM network with an encoder-decoder structure is trained using the station-observed meteorological elements, numerical weather prediction model-forecasted radiance data, numerical weather prediction model-forecasted meteorological elements, various intrinsic mode functions, and numerical weather prediction model-forecasted radiance data at the correction time as input, and the station-observed radiance data on the correction date as output. The model is then validated using a validation set to obtain the irradiance numerical model forecast correction model, which includes: The station observation meteorological elements, numerical weather prediction model forecast radiation data, numerical weather prediction model forecast meteorological elements, and each intrinsic mode function are input into the CNN structure of the encoder at each time step in the encoder time sequence input length before correction time in the training set to obtain the encoder feature vector at each time step. The current encoder feature vector and the cell state and hidden state output by the encoder's LSTM structure in the previous time step are input into the LSTM structure to obtain the cell state and hidden state output by the LSTM structure in the current time step. This process is repeated to obtain the cell state and hidden state output in the time step before the correction date. The cell state and hidden state output by the encoder in the previous time step are preset values. The numerical weather prediction model forecast radiation data, numerical weather prediction model forecast meteorological elements, and the correction result of the previous time step are input into the CNN structure of the decoder at each time step starting from the correction time in the training set to obtain the decoder feature vector at each time step. The current decoder feature vector and the cell state and hidden state output by the decoder's LSTM structure at the previous time step are input into the LSTM structure to obtain the cell state and hidden state output by the LSTM structure at the current time step. The hidden state output by the LSTM structure at the current time step is input into the fully connected layer to obtain the correction result at the current time step. This process is repeated continuously to obtain the correction results at each time step in the time sequence of the decoder starting from the correction time step. The cell state and hidden state output by the decoder at the previous time step are the cell state and hidden state output by the encoder at the time step before the correction date. The prediction error is obtained by subtracting the correction result output by the decoder from the corresponding field observation radiation data. The parameters of the encoder and decoder are adjusted based on the error until the model passes the validation set, thus obtaining the irradiance numerical model forecast correction model.

[0008] Preferably, the meteorological elements include one or more of the following: temperature, humidity, or surface wind field component.

[0009] Preferably, the acquisition of station-observed radiation data, station-observed meteorological elements, numerical weather prediction model forecast radiation data, and numerical weather prediction model forecast meteorological elements for a set period prior to the correction time, as well as numerical weather prediction model forecast radiation data and numerical weather prediction model forecast meteorological elements on the correction date, includes: Collect on-site radiation data and on-site meteorological data of new energy power stations within the study area for a set period before the correction time; Select a suitable physical parameterization scheme for the WRF-solar model to conduct irradiance simulation experiments, and at the same time conduct dynamic downscaling experiments to output gridded forecast radiation data and gridded forecast meteorological elements with a set duration and correction date before the correction time. The gridded forecast radiation data and gridded forecast meteorological elements were standardized using a batch normalization method. Then, the gridded forecast radiation data and gridded forecast meteorological elements were interpolated to the station locations using a bilinear interpolation method to obtain numerical weather prediction model forecast radiation data and numerical weather prediction model forecast meteorological elements that match the stations.

[0010] Based on the same inventive concept, the present invention also provides a deep learning irradiance numerical pattern forecasting correction system that incorporates clustering, including: a data acquisition module, a type classification module, a mode decomposition module and a irradiance correction module; The data acquisition module is used to acquire the station observation radiation data, station observation meteorological elements, numerical weather prediction model forecast radiation data and numerical weather prediction model forecast meteorological elements for a set period before the correction time, as well as the numerical weather prediction model forecast radiation data and numerical weather prediction model forecast meteorological elements on the correction date. The type classification module is used to cluster the station observation radiation data and station observation meteorological elements using a clustering algorithm to classify multiple weather types. The mode decomposition module is used to process the field observation radiation data using a variational mode decomposition algorithm to obtain multiple intrinsic mode functions; The radiation correction module is used to input the station observation meteorological elements, numerical weather prediction model forecast radiation data, numerical weather prediction model forecast meteorological elements and various intrinsic mode functions, as well as the numerical weather prediction model forecast radiation data and numerical weather prediction model forecast meteorological elements on the correction date, into the pre-trained irradiance numerical model forecast correction model based on the weather type, so as to obtain the numerical weather prediction model forecast radiation correction result on the correction date. The irradiance numerical model forecast correction model is a CNN-LSTM network with an encoder-decoder structure. The irradiance numerical model forecast correction model is trained using historical field observation irradiance data, field observation meteorological data, numerical weather prediction model forecast irradiance data, and numerical weather prediction model forecast meteorological elements.

[0011] Preferably, the training process of the irradiance numerical model prediction correction model in the radiation correction module includes: Acquire historical data on field observations of radiation, meteorological elements observed at the field, radiation data and meteorological elements predicted by numerical weather prediction models, and divide them into training and validation sets according to a set ratio. Clustering algorithms were used to cluster the observed radiation data and meteorological elements from the stations, and to classify them into multiple weather types. The field observation radiation data were processed using a variational mode decomposition algorithm to obtain multiple intrinsic mode functions; For each weather type, a CNN-LSTM network with an encoder-decoder structure is trained using the following inputs: meteorological elements observed at the station for a set duration before the correction time in the training set, radiometric data predicted by the numerical weather prediction model, meteorological elements predicted by the numerical weather prediction model, various intrinsic mode functions, and radiometric data predicted by the numerical weather prediction model at the correction time. The output is the radiometric data observed at the station on the correction date. The model is then validated using a validation set to obtain the irradiance numerical model forecast correction model.

[0012] Preferably, the radiation correction module takes as input station-observed meteorological elements, numerical weather prediction model-forecasted radiation data, numerical weather prediction model-forecasted meteorological elements, various intrinsic mode functions, and numerical weather prediction model-forecasted radiation data at the correction time, and as output station-observed radiation data on the correction date, a CNN-LSTM network with an encoder-decoder structure is trained, and validated using a validation set to obtain an irradiance numerical model forecast correction model, including: The station observation meteorological elements, numerical weather prediction model forecast radiation data, numerical weather prediction model forecast meteorological elements, and each intrinsic mode function are input into the CNN structure of the encoder at each time step in the encoder time sequence input length before correction time in the training set to obtain the encoder feature vector at each time step. The current encoder feature vector and the cell state and hidden state output by the encoder's LSTM structure in the previous time step are input into the LSTM structure to obtain the cell state and hidden state output by the LSTM structure in the current time step. This process is repeated to obtain the cell state and hidden state output in the time step before the correction date. The cell state and hidden state output by the encoder in the previous time step are preset values. The numerical weather prediction model forecast radiation data, numerical weather prediction model forecast meteorological elements, and the correction result of the previous time step are input into the CNN structure of the decoder at each time step starting from the correction time in the training set to obtain the decoder feature vector at each time step. The current decoder feature vector and the cell state and hidden state output by the decoder's LSTM structure at the previous time step are input into the LSTM structure to obtain the cell state and hidden state output by the LSTM structure at the current time step. The hidden state output by the LSTM structure at the current time step is input into the fully connected layer to obtain the correction result at the current time step. This process is repeated continuously to obtain the correction results at each time step in the time sequence of the decoder starting from the correction time step. The cell state and hidden state output by the decoder at the previous time step are the cell state and hidden state output by the encoder at the time step before the correction date. The prediction error is obtained by subtracting the correction result output by the decoder from the corresponding field observation radiation data. The parameters of the encoder and decoder are adjusted based on the error until the model passes the validation set, thus obtaining the irradiance numerical model forecast correction model.

[0013] Preferably, the meteorological elements include one or more of the following: temperature, humidity, or surface wind field component.

[0014] Preferably, the data acquisition module is specifically used for: Collect on-site radiation data and on-site meteorological data of new energy power stations within the study area for a set period before the correction time; Select a suitable physical parameterization scheme for the WRF-solar model to conduct irradiance simulation experiments, and at the same time conduct dynamic downscaling experiments to output gridded forecast radiation data and gridded forecast meteorological elements with a set duration and correction date before the correction time. The gridded forecast radiation data and gridded forecast meteorological elements were standardized using a batch normalization method. Then, the gridded forecast radiation data and gridded forecast meteorological elements were interpolated to the station locations using a bilinear interpolation method to obtain numerical weather prediction model forecast radiation data and numerical weather prediction model forecast meteorological elements that match the stations.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention provides a deep learning-based method and system for correcting numerical irradiance model forecasts by incorporating clustering. The method includes: acquiring station-observed irradiance data, station-observed meteorological elements, numerical weather prediction model (NTM) forecast irradiance data, and NTM forecast meteorological elements for a predetermined duration before the correction date, as well as NTM forecast irradiance data and NTM forecast meteorological elements on the correction date; using a clustering algorithm to cluster the station-observed irradiance data and station-observed meteorological elements to classify multiple weather types; using a variational mode decomposition (VMD) algorithm to process the station-observed irradiance data to obtain multiple intrinsic mode functions (IMFs); and based on the weather type, classifying the station-observed meteorological elements for a predetermined duration before the correction date... The present invention incorporates numerical weather prediction (NTM) data on irradiance, meteorological elements, and intrinsic mode functions (IMFs) from a numerical weather prediction model, along with NTM data and meteorological elements for the correction date, into a pre-trained numerical irradiance model. This yields the corrected irradiance forecast from the NTM for the correction date. The irradiance model is a CNN-LSTM network with an encoder-decoder structure, trained using historical field observation irradiance data, field observation meteorological elements, and NTM data. The invention introduces a clustering algorithm to group information sequences under the same generalized weather type into the same cluster. A variational mode decomposition algorithm decomposes historical time-series signals into various IMFs, removing redundant information and obtaining effective features. These optimizations enhance the feature learning task, enabling more effective deep learning model training and error correction. Attached Figure Description

[0016] Figure 1 A flowchart of a deep learning method for correcting numerical irradiance pattern forecasts by incorporating clustering, provided by this invention; Figure 2 Comparison chart of correction effects provided by the present invention; Figure 3A flowchart illustrating a specific example of a deep learning-based numerical irradiance pattern forecast correction method incorporating clustering, provided by this invention; Figure 4 The encoder-decoder structure based on CNN-LSTM provided by this invention; Figure 5 The data input diagram of the encoder-decoder provided by this invention; Figure 6 This is a schematic diagram of a deep learning-based numerical irradiance model forecasting correction system that incorporates clustering, as provided by the present invention. Detailed Implementation

[0017] Irradiance is the core of photovoltaic power generation forecasting, and different weather factors have significantly different impacts on irradiance. Improving forecast accuracy by correcting biases in numerical forecast data is an important research topic in meteorological forecasting. This invention proposes a deep learning-based method and system for correcting numerical irradiance forecasts using clustering. Its purpose is to revise forecast data based on existing historical measured data from photovoltaic power plants and station data provided by meteorological stations as primary data for irradiance forecasting research. Stable and accurate irradiance forecasts are necessary to ensure accurate power generation, thus contributing to the stability of photovoltaic grid-connected power and playing a crucial role in the safe dispatch and scientific planning of the power system.

[0018] Example 1: A deep learning-based method for correcting numerical irradiance model forecasts by incorporating clustering, such as... Figure 1 As shown, it includes: Step 1: Obtain the station observation radiation data, station observation meteorological elements, numerical weather prediction model forecast radiation data and numerical weather prediction model forecast meteorological elements for the set duration before the correction time, as well as the numerical weather prediction model forecast radiation data and numerical weather prediction model forecast meteorological elements on the correction date. Step 2: Use clustering algorithms to cluster the station-observed radiation data and station-observed meteorological elements to classify multiple weather types; Step 3: Use the variational mode decomposition algorithm to process the field observation radiation data to obtain multiple intrinsic mode functions; Step 4: Based on the weather type, input the station observation meteorological elements, numerical weather prediction model forecast radiation data, numerical weather prediction model forecast meteorological elements and various intrinsic mode functions for the set duration before the correction time, as well as the numerical weather prediction model forecast radiation data and numerical weather prediction model forecast meteorological elements on the correction date, into the pre-trained irradiance numerical model forecast correction model to obtain the numerical weather prediction model forecast radiation correction result on the correction date; Among them, the irradiance numerical model forecast correction model is a CNN-LSTM network with an encoder-decoder structure. The irradiance numerical model forecast correction model is trained using historical station observation irradiance data, station observation meteorological data, numerical weather prediction model forecast irradiance data, and numerical weather prediction model forecast meteorological elements.

[0019] Before step 1, the process of training the numerical irradiance model forecast correction model includes: (1) Select the study area and confirm the time series length of historical data, obtain numerical model data and observation data of the study area, select a suitable physical parameterization scheme for NWP numerical model forecasting, and perform dynamic downscaling to make the grid in the study area more refined and the simulation data more accurate.

[0020] (2) The bilinear interpolation algorithm was used to interpolate the numerical model forecast data and the selected meteorological feature to the ground radiation station. The data preprocessing work, such as spatiotemporal calibration, data cleaning, outlier and missing value handling, and data standardization, was carried out to construct a historical dataset containing ground station radiation, meteorological data and model forecast radiation, and meteorological data. At the same time, the training set and validation set were divided in an 8:2 ratio according to the length of the time series of the historical dataset.

[0021] (3) The fluctuation characteristics of solar irradiance vary under different weather conditions. Based on clustering algorithms, such as Kmeans++ clustering algorithm, the historical dataset is divided into three categories of generalized weather according to the radiation data of ground measured stations in the historical dataset before the correction date. Finally, it is actually divided into three categories of generalized weather: sunny, cloudy, and overcast.

[0022] (4) Treat each day's data as a data object, which is the radiation data observed at the station ( ), meteorological elements observed at the station ( ), NWP (numerical weather prediction) model forecast radiation data ( ) and NWP model forecast meteorological elements ( The meteorological elements observed at the field and those predicted by NWP include temperature, humidity, and 10m UV wind. The VMD variational mode decomposition algorithm is used to decompose the time series signal of historical field observation radiation data into various intrinsic mode functions, removing redundant information and obtaining effective features. This invention fixes a one-year time window and slides the time window during each decomposition to obtain new input data. This is to ensure that the data is obtained based on historical data prior to the correction time and to prevent information leakage.

[0023] (5) Establish a CNN-LSTM network with an encoder-decoder structure. The VMD decomposition results, along with historical meteorological data from station observations, model-predicted irradiance data, and model-predicted meteorological data, are used as inputs to the encoder-decoder. Training and modeling are performed separately based on the training set, and parameter tuning is performed based on the validation set before finalization.

[0024] After the numerical irradiance model forecast correction model is trained, the deep learning numerical irradiance model forecast correction process, i.e., steps 1-4, is as follows. Since the K-means++ clustering algorithm for classifying generalized weather types and the VMD algorithm for decomposition are the same as the training process, they will not be described again.

[0025] (6) The historical data of 7 days before the date to be corrected (i.e. the set duration) and the data of one day before the date to be corrected are formed into a time series dataset according to the requirements of the model input data. After the generalized weather type is divided by the Kmeans++ clustering algorithm, the VMD algorithm is decomposed, and the data is assigned to the encoder-decoder structure, the irradiance revision result of the model forecast data for the correction date is output.

[0026] Error analysis reveals significant differences in the accuracy of irradiance model forecasts under different weather conditions, and most existing correction methods do not consider modeling separately for different meteorological conditions. This invention introduces the K-means++ clustering algorithm to group information sequences under the same generalized weather type into the same cluster; the VMD variational mode decomposition algorithm can decompose historical time series signals into various intrinsic mode functions, removing redundant information and obtaining effective features; these optimize the feature learning task, enabling more effective deep learning model training and error correction. Figure 2 As shown, the error correction results of XGBoost (left) and the method of this invention (right) were tested using the same dataset. Figure 2In the diagram, the ordinate Observed GHI represents the measured global horizontal irradiance, i.e., the irradiance data actually measured by ground stations. The abscissa Estimated GHI represents the corrected global horizontal irradiance, i.e., the irradiance results output by XGBoost (left) and the method of this invention (right). R represents the coefficient of determination, used to measure the goodness of fit between the model estimate and the measured value. The closer R is to 1, the better the model's prediction effect. RMSE represents the root mean square error, used to measure the magnitude of the error between the model estimate and the measured value. The smaller the RMSE, the higher the model accuracy. Y=ax+b: represents the linear fitting equation between the measured value and the estimated value, where y corresponds to "Observed GHI" and x corresponds to "Estimated GHI"; N represents the sample size. The results verify that the proposed method can effectively improve the correlation coefficient between the irradiance of the model forecast data and the actual irradiance, and has a better correction effect than traditional machine learning methods.

[0027] Example 2: The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0028] like Figure 3 As shown, the flowchart of a specific example of a deep learning-based numerical irradiance pattern forecast correction method incorporating clustering is as follows: Step 1: Data Collection. The study area is set as Jiangsu Province. First, four years of historical radiometer data from new energy power plants within the study area, along with corresponding meteorological data, are collected. A suitable physical parameterization scheme for the WRF-solar model is selected for irradiance simulation experiments, including microphysical process parameters, longwave radiation schemes, cumulus parameterization schemes, and land surface process schemes. The shortwave radiation scheme uses the RRTMG scheme, and dynamic downscaling experiments are conducted simultaneously. Output factors include ambient temperature, humidity, shortwave radiative flux, and surface wind field components.

[0029] Step two involves standardizing the data using Batch Normalization (BN). Bilinear interpolation is used to interpolate the model grid data to the station locations, ensuring a match with the observed data from the meteorological stations. If a sample has no model forecast or observational data or is invalid, that sample is discarded. The training and validation sets are divided in an 8:2 ratio, with the historical data from the previous three years used as the training set and the data from the following year used as the validation set.

[0030] Step 3 uses historical measured irradiance, temperature, humidity, 10 m wind field components (u, v), theoretical clear-sky irradiance, and solar zenith angle sequences as inputs. Time-series statistics are constructed for each variable, and features reflecting weather conditions, such as gradient changes and stability indices, are extracted from the irradiance sequence. The number of clusters is determined to be 3, and the K-means++ clustering algorithm is used to cluster the data, corresponding to three generalized weather types: sunny, cloudy, and overcast. Subsequent model training and corrections are performed for each weather type.

[0031] Step four: Treat each day's data as a data object, derived from the station's observed radiation data ( ), meteorological elements observed at the station ( ), NWP model forecast radiation data ( ) and NWP model forecast meteorological elements ( This invention uses the VMD variational mode decomposition algorithm to decompose the time series signals of historical field observation radiation data into various intrinsic mode functions, removing redundant information and obtaining effective features. The invention fixes a one-year time window and slides the time window during each decomposition to obtain seven new sub-signals. This is to ensure that the sub-signals are obtained based on historical data prior to the correction time and to prevent information leakage.

[0032] Step 5: Build a CNN-LSTM network with an encoder-decoder structure. For example... Figure 4 As shown, and h ... and The final cell state and hidden state are fed into the first LSTM unit of the decoder.

[0033] For the decoder, and This represents the initial state of the first LSTM unit. At time k, the input to the decoder is... and the correction results at time k-1. Composed of data objects After passing through CNN, LSTM, and fully connected layers ( Figure 4 (LR) and the output of the LSTM layer and The k+1 time step is input together, and the correction result is obtained by iterative iteration. A CNN-LSTM network with an encoder-decoder structure is used to model the model based on the training set, and the model is finalized after parameter tuning based on the validation set, resulting in the irradiance numerical pattern forecast correction model.

[0034] The data inputs of the encoder and decoder are as follows Figure 5 As shown.

[0035] First, the continuous historical time-series data is constructed into a fixed-length time window. All features at each time step within the window are combined to form a data object. l and h are the temporal input lengths of the encoder and decoder, respectively. For example, at historical time tl, To analyze the observed radiation data The VMD decomposition yields m sub-components. , , These are, respectively, the field observation radiation data, field observation meteorological elements, NWP model forecast radiation data, and NWP model forecast meteorological elements at that moment. Therefore, It is the basic unit that constitutes a time window, and its sequence Together, they serve as input to the encoder, used to extract historical information and generate a compressed temporal representation.

[0036] For the future prediction phase, the decoder constructs a data object at each prediction moment to form a data object. For example, at time t+1, , , Composed of data objects This is the input to the decoder, where This is the correction result for the previous time step t. The final cell state output by the encoder after processing the history window. With hidden state Used as the initial state for the decoder LSTM unit, followed by the decoder dependency sequence. Gradually generate future correction results.

[0037] Step six involves creating a time-series dataset from the historical data of the seven days prior to the correction date and the data for the day of the correction date, according to the requirements of the model input data. This dataset is then processed using the K-means++ clustering algorithm to classify generalized weather types and the VMD algorithm for decomposition. The corresponding correction model is loaded, and the data is assigned to the encoder-decoder structure. Finally, the irradiance revision results for the model forecast data on the correction date are output.

[0038] The effects achieved by this application are summarized as follows: (1) Considering the significant differences in the fluctuation characteristics of solar irradiance under different weather conditions, this invention first introduces the Kmeans++ clustering algorithm, using historical measured irradiance, temperature, humidity, 10m UV wind, theoretical clear-sky irradiance, and solar zenith angle as inputs. Different weather patterns are distinguished based on the morphology of the model input data sequence, classifying the data into three generalized weather types: sunny, cloudy, and overcast. Correction models are established for different weather types to achieve classification modeling and type-based correction, thereby fully utilizing the regular characteristics under different weather conditions and effectively avoiding the problems of error amplification and insufficient adaptability that occur when using a single model to process all weather data in existing methods.

[0039] (2) To address the potential noise and redundancy in irradiance time series data, this invention employs the VMD (Variational Mode Decomposition) algorithm to decompose historical station observation irradiance data. This process separates the complex, non-stationary original signal into several Intrinsic Mode Functions (IMFs), removes redundant information, and extracts the most useful key feature components for correction. Compared to most existing methods that directly utilize the original time series data for modeling, VMD can effectively reduce the impact of data noise on model training, improve feature representation capabilities, and provide higher-quality input features for subsequent deep learning modeling.

[0040] (3) To fully explore the complex nonlinear relationships in meteorological elements, numerical forecast data, and observational data, this invention constructs an encoder-decoder deep learning network structure based on CNN-LSTM. The convolutional neural network (CNN) can extract local spatial features from multidimensional time-series data, while the long short-term memory network (LSTM) can capture long-term dependencies in the time series. The combination of these two forms an end-to-end feature learning and prediction framework. Compared to traditional machine learning methods (such as XGBoost), this structure is better able to handle large-scale, multidimensional data with strong spatiotemporal correlations, improving the accuracy and stability of irradiance correction.

[0041] (4) By using clustering, VMD feature extraction, and deep learning technologies in a coordinated manner, an end-to-end data correction scheme for historical irradiance observation data of stations was created for different weather categories. This scheme breaks through the limitations of a single technology, can flexibly adjust the model strategy for different meteorological conditions, and adaptively extract effective features through a data-driven approach, forming a multi-level, multi-stage collaborative optimization correction process.

[0042] Example 3: Based on the same inventive concept, this invention also provides a deep learning-based numerical irradiance pattern prediction correction system that incorporates clustering, such as... Figure 6 As shown, it includes: a data acquisition module, a type classification module, a mode decomposition module, and a radiation correction module; The data acquisition module is used to acquire the station observation radiation data, station observation meteorological elements, numerical weather prediction model forecast radiation data and numerical weather prediction model forecast meteorological elements for a set period before the correction time, as well as the numerical weather prediction model forecast radiation data and numerical weather prediction model forecast meteorological elements on the correction date. The type classification module is used to cluster the station observation radiation data and station observation meteorological elements using a clustering algorithm to classify multiple weather types. The mode decomposition module is used to process the field observation radiation data using a variational mode decomposition algorithm to obtain multiple intrinsic mode functions; The radiation correction module is used to input the station observation meteorological elements, numerical weather prediction model forecast radiation data, numerical weather prediction model forecast meteorological elements and various intrinsic mode functions, as well as the numerical weather prediction model forecast radiation data and numerical weather prediction model forecast meteorological elements on the correction date, into the pre-trained irradiance numerical model forecast correction model based on the weather type, so as to obtain the numerical weather prediction model forecast radiation correction result on the correction date. The irradiance numerical model forecast correction model is a CNN-LSTM network with an encoder-decoder structure. The irradiance numerical model forecast correction model is trained using historical field observation irradiance data, field observation meteorological data, numerical weather prediction model forecast irradiance data, and numerical weather prediction model forecast meteorological elements.

[0043] Preferably, the training process of the irradiance numerical model prediction correction model in the radiation correction module includes: Acquire historical data on field observations of radiation, meteorological elements observed at the field, radiation data and meteorological elements predicted by numerical weather prediction models, and divide them into training and validation sets according to a set ratio. Clustering algorithms were used to cluster the observed radiation data and meteorological elements from the stations, and to classify them into multiple weather types. The field observation radiation data were processed using a variational mode decomposition algorithm to obtain multiple intrinsic mode functions; For each weather type, a CNN-LSTM network with an encoder-decoder structure is trained using the following inputs: meteorological elements observed at the station for a set duration before the correction time in the training set, radiometric data predicted by the numerical weather prediction model, meteorological elements predicted by the numerical weather prediction model, various intrinsic mode functions, and radiometric data predicted by the numerical weather prediction model at the correction time. The output is the radiometric data observed at the station on the correction date. The model is then validated using a validation set to obtain the irradiance numerical model forecast correction model.

[0044] Preferably, the radiation correction module takes as input station-observed meteorological elements, numerical weather prediction model-forecasted radiation data, numerical weather prediction model-forecasted meteorological elements, various intrinsic mode functions, and numerical weather prediction model-forecasted radiation data at the correction time, and as output station-observed radiation data on the correction date, a CNN-LSTM network with an encoder-decoder structure is trained, and validated using a validation set to obtain an irradiance numerical model forecast correction model, including: The station observation meteorological elements, numerical weather prediction model forecast radiation data, numerical weather prediction model forecast meteorological elements, and each intrinsic mode function are input into the CNN structure of the encoder at each time step in the encoder time sequence input length before correction time in the training set to obtain the encoder feature vector at each time step. The current encoder feature vector and the cell state and hidden state output by the encoder's LSTM structure in the previous time step are input into the LSTM structure to obtain the cell state and hidden state output by the LSTM structure in the current time step. This process is repeated to obtain the cell state and hidden state output in the time step before the correction date. The cell state and hidden state output by the encoder in the previous time step are preset values. The numerical weather prediction model forecast radiation data, numerical weather prediction model forecast meteorological elements, and the correction result of the previous time step are input into the CNN structure of the decoder at each time step starting from the correction time in the training set to obtain the decoder feature vector at each time step. The current decoder feature vector and the cell state and hidden state output by the decoder's LSTM structure at the previous time step are input into the LSTM structure to obtain the cell state and hidden state output by the LSTM structure at the current time step. The hidden state output by the LSTM structure at the current time step is input into the fully connected layer to obtain the correction result at the current time step. This process is repeated continuously to obtain the correction results at each time step in the time sequence of the decoder starting from the correction time step. The cell state and hidden state output by the decoder at the previous time step are the cell state and hidden state output by the encoder at the time step before the correction date. The prediction error is obtained by subtracting the correction result output by the decoder from the corresponding field observation radiation data. The parameters of the encoder and decoder are adjusted based on the error until the model passes the validation set, thus obtaining the irradiance numerical model forecast correction model.

[0045] Preferably, the meteorological elements include one or more of the following: temperature, humidity, or surface wind field component.

[0046] Preferably, the data acquisition module is specifically used for: Collect on-site radiation data and on-site meteorological data of new energy power stations within the study area for a set period before the correction time; Select a suitable physical parameterization scheme for the WRF-solar model to conduct irradiance simulation experiments, and at the same time conduct dynamic downscaling experiments to output gridded forecast radiation data and gridded forecast meteorological elements with a set duration and correction date before the correction time. The gridded forecast radiation data and gridded forecast meteorological elements were standardized using a batch normalization method. Then, the gridded forecast radiation data and gridded forecast meteorological elements were interpolated to the station locations using a bilinear interpolation method to obtain numerical weather prediction model forecast radiation data and numerical weather prediction model forecast meteorological elements that match the stations.

[0047] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0048] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0049] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0050] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0051] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of the claims of the present invention pending approval.

Claims

1. A deep learning-based method for correcting numerical irradiance model forecasts by incorporating clustering, characterized in that, include: Obtain the station observation radiation data, station observation meteorological elements, numerical weather prediction model forecast radiation data and numerical weather prediction model forecast meteorological elements for a set period before the correction time, as well as the numerical weather prediction model forecast radiation data and numerical weather prediction model forecast meteorological elements on the correction date. Clustering algorithms were used to cluster the observed radiation data and meteorological elements from the stations, and to classify them into multiple weather types. The field observation radiation data were processed using a variational mode decomposition algorithm to obtain multiple intrinsic mode functions; Based on the weather type, the station observation meteorological elements, numerical weather prediction model forecast radiation data, numerical weather prediction model forecast meteorological elements and various intrinsic mode functions, as well as the numerical weather prediction model forecast radiation data and numerical weather prediction model forecast meteorological elements on the correction date, are input into the pre-trained irradiance numerical model forecast correction model to obtain the numerical weather prediction model forecast radiation correction result on the correction date. The irradiance numerical model forecast correction model is a CNN-LSTM network with an encoder-decoder structure. The irradiance numerical model forecast correction model is trained using historical field observation irradiance data, field observation meteorological data, numerical weather prediction model forecast irradiance data, and numerical weather prediction model forecast meteorological elements.

2. The method as described in claim 1, characterized in that, The training process of the numerical irradiance model prediction correction model includes: Acquire historical data on field observations of radiation, meteorological elements observed at the field, radiation data and meteorological elements predicted by numerical weather prediction models, and divide them into training and validation sets according to a set ratio. Clustering algorithms were used to cluster the observed radiation data and meteorological elements from the stations, and to classify them into multiple weather types. The field observation radiation data were processed using a variational mode decomposition algorithm to obtain multiple intrinsic mode functions; For each weather type, a CNN-LSTM network with an encoder-decoder structure is trained using the following inputs: meteorological elements observed at the station for a set duration before the correction time in the training set, radiometric data predicted by the numerical weather prediction model, meteorological elements predicted by the numerical weather prediction model, various intrinsic mode functions, and radiometric data predicted by the numerical weather prediction model at the correction time. The output is the radiometric data observed at the station on the correction date. The model is then validated using a validation set to obtain the irradiance numerical model forecast correction model.

3. The method as described in claim 2, characterized in that, The CNN-LSTM network with an encoder-decoder structure is trained using the station-observed meteorological elements, numerical weather prediction model (NTM) forecast radiance data, NTM forecast meteorological elements, various intrinsic mode functions, and NTM forecast radiance data at the correction time as inputs, and the station-observed radiance data on the correction date as outputs. The model is then validated using a validation set to obtain the irradiance numerical model forecast correction model, which includes: The station observation meteorological elements, numerical weather prediction model forecast radiation data, numerical weather prediction model forecast meteorological elements, and each intrinsic mode function are input into the CNN structure of the encoder at each time step in the encoder time sequence input length before correction time in the training set to obtain the encoder feature vector at each time step. The current encoder feature vector and the cell state and hidden state output by the encoder's LSTM structure in the previous time step are input into the LSTM structure to obtain the cell state and hidden state output by the LSTM structure in the current time step. This process is repeated to obtain the cell state and hidden state output in the time step before the correction date. The cell state and hidden state output by the encoder in the previous time step are preset values. The numerical weather prediction model forecast radiation data, numerical weather prediction model forecast meteorological elements, and the correction result of the previous time step are input into the CNN structure of the decoder at each time step starting from the correction time in the training set to obtain the decoder feature vector at each time step. The current decoder feature vector and the cell state and hidden state output by the decoder's LSTM structure at the previous time step are input into the LSTM structure to obtain the cell state and hidden state output by the LSTM structure at the current time step. The hidden state output by the LSTM structure at the current time step is input into the fully connected layer to obtain the correction result at the current time step. This process is repeated continuously to obtain the correction results at each time step in the time sequence of the decoder starting from the correction time step. The cell state and hidden state output by the decoder at the previous time step are the cell state and hidden state output by the encoder at the time step before the correction date. The prediction error is obtained by subtracting the correction result output by the decoder from the corresponding field observation radiation data. The parameters of the encoder and decoder are adjusted based on the error until the model passes the validation set, thus obtaining the irradiance numerical model forecast correction model.

4. The method as described in claim 1, characterized in that, The meteorological elements include one or more of the following: temperature, humidity, or surface wind field components.

5. The method as described in claim 1, characterized in that, The acquisition of station-observed radiation data, station-observed meteorological elements, numerical weather prediction model forecasted radiation data, and numerical weather prediction model forecasted meteorological elements for a set duration prior to the correction time, as well as numerical weather prediction model forecasted radiation data and numerical weather prediction model forecasted meteorological elements on the correction date, includes: Collect on-site radiation data and on-site meteorological data of new energy power stations within the study area for a set period before the correction time; Select a suitable physical parameterization scheme for the WRF-solar model to conduct irradiance simulation experiments, and at the same time conduct dynamic downscaling experiments to output gridded forecast radiation data and gridded forecast meteorological elements with a set duration and correction date before the correction time. The gridded forecast radiation data and gridded forecast meteorological elements were standardized using a batch normalization method. Then, the gridded forecast radiation data and gridded forecast meteorological elements were interpolated to the station locations using a bilinear interpolation method to obtain numerical weather prediction model forecast radiation data and numerical weather prediction model forecast meteorological elements that match the stations.

6. A deep learning-based numerical irradiance model forecasting correction system incorporating clustering, characterized in that, include: Data acquisition module, type classification module, mode decomposition module, and radiation correction module; The data acquisition module is used to acquire the station observation radiation data, station observation meteorological elements, numerical weather prediction model forecast radiation data and numerical weather prediction model forecast meteorological elements for a set period before the correction time, as well as the numerical weather prediction model forecast radiation data and numerical weather prediction model forecast meteorological elements on the correction date. The type classification module is used to cluster the station observation radiation data and station observation meteorological elements using a clustering algorithm to classify multiple weather types. The mode decomposition module is used to process the field observation radiation data using a variational mode decomposition algorithm to obtain multiple intrinsic mode functions; The radiation correction module is used to input the station observation meteorological elements, numerical weather prediction model forecast radiation data, numerical weather prediction model forecast meteorological elements and various intrinsic mode functions, as well as the numerical weather prediction model forecast radiation data and numerical weather prediction model forecast meteorological elements on the correction date, into the pre-trained irradiance numerical model forecast correction model based on the weather type, so as to obtain the numerical weather prediction model forecast radiation correction result on the correction date. The irradiance numerical model forecast correction model is a CNN-LSTM network with an encoder-decoder structure. The irradiance numerical model forecast correction model is trained using historical field observation irradiance data, field observation meteorological data, numerical weather prediction model forecast irradiance data, and numerical weather prediction model forecast meteorological elements.

7. The system as described in claim 6, characterized in that, The training process of the irradiance numerical model prediction correction model in the radiation correction module includes: Acquire historical data on field observations of radiation, meteorological elements observed at the field, radiation data and meteorological elements predicted by numerical weather prediction models, and divide them into training and validation sets according to a set ratio. Clustering algorithms were used to cluster the observed radiation data and meteorological elements from the stations, and to classify them into multiple weather types. The field observation radiation data were processed using a variational mode decomposition algorithm to obtain multiple intrinsic mode functions; For each weather type, a CNN-LSTM network with an encoder-decoder structure is trained using the following inputs: meteorological elements observed at the station for a set duration before the correction time in the training set, radiometric data predicted by the numerical weather prediction model, meteorological elements predicted by the numerical weather prediction model, various intrinsic mode functions, and radiometric data predicted by the numerical weather prediction model at the correction time. The output is the radiometric data observed at the station on the correction date. The model is then validated using a validation set to obtain the irradiance numerical model forecast correction model.

8. The system as described in claim 7, characterized in that, The radiation correction module takes as input station-observed meteorological elements, numerical weather prediction model (NTM) forecast radiation data, NTM forecast meteorological elements, various intrinsic mode functions, and NTM forecast radiation data at the correction time, and uses station-observed radiation data on the correction date as output to train a CNN-LSTM network with an encoder-decoder structure. The network is then validated using a validation set to obtain the irradiance numerical model forecast correction model, which includes: The station observation meteorological elements, numerical weather prediction model forecast radiation data, numerical weather prediction model forecast meteorological elements, and each intrinsic mode function are input into the CNN structure of the encoder at each time step in the encoder time sequence input length before correction time in the training set to obtain the encoder feature vector at each time step. The current encoder feature vector and the cell state and hidden state output by the encoder's LSTM structure in the previous time step are input into the LSTM structure to obtain the cell state and hidden state output by the LSTM structure in the current time step. This process is repeated to obtain the cell state and hidden state output in the time step before the correction date. The cell state and hidden state output by the encoder in the previous time step are preset values. The numerical weather prediction model forecast radiation data, numerical weather prediction model forecast meteorological elements, and the correction result of the previous time step are input into the CNN structure of the decoder at each time step starting from the correction time in the training set to obtain the decoder feature vector at each time step. The current decoder feature vector and the cell state and hidden state output by the decoder's LSTM structure at the previous time step are input into the LSTM structure to obtain the cell state and hidden state output by the LSTM structure at the current time step. The hidden state output by the LSTM structure at the current time step is input into the fully connected layer to obtain the correction result at the current time step. This process is repeated continuously to obtain the correction results at each time step in the time sequence of the decoder starting from the correction time step. The cell state and hidden state output by the decoder at the previous time step are the cell state and hidden state output by the encoder at the time step before the correction date. The prediction error is obtained by subtracting the correction result output by the decoder from the corresponding field observation radiation data. The parameters of the encoder and decoder are adjusted based on the error until the model passes the validation set, thus obtaining the irradiance numerical model forecast correction model.

9. The system as described in claim 6, characterized in that, The meteorological elements include one or more of the following: temperature, humidity, or surface wind field components.

10. The system as described in claim 6, characterized in that, The data acquisition module is specifically used for: Collect on-site radiation data and on-site meteorological data of new energy power stations within the study area for a set period before the correction time; Select a suitable physical parameterization scheme for the WRF-solar model to conduct irradiance simulation experiments, and at the same time conduct dynamic downscaling experiments to output gridded forecast radiation data and gridded forecast meteorological elements with a set duration and correction date before the correction time. The gridded forecast radiation data and gridded forecast meteorological elements were standardized using a batch normalization method. Then, the gridded forecast radiation data and gridded forecast meteorological elements were interpolated to the station locations using a bilinear interpolation method to obtain numerical weather prediction model forecast radiation data and numerical weather prediction model forecast meteorological elements that match the stations.

Images