Ultra-short-term photovoltaic power prediction method, system and equipment based on multivariate data fusion and medium

Abstract

Aiming at the problem of low photovoltaic power prediction accuracy, the invention provides an ultra-short-term photovoltaic power prediction method, system, equipment and medium based on multivariate data fusion, and the method comprises the steps: inputting underlying surface data, ground meteorological monitoring data and satellite monitoring data into a WRF-Solar model, and predicting solar irradiance; and inputting the solar irradiance, the historical weather data and the historical data of the photovoltaic power station into the LSTM power prediction model, and predicting the ultra-short-time photovoltaic power. By fusing multi-source heterogeneous data such as underlying surface data, ground real-time meteorological monitoring data, satellite monitoring data and photovoltaic module operation states and combining collaborative optimization of a machine learning model and a physical model, the ultra-short-term numerical weather forecasting precision can be effectively improved, short-term forecasting errors in extreme weather and complex scenes are reduced, and the forecasting precision of the ultra-short-term numerical weather can be effectively improved. Therefore, the accuracy of ultra-short-term photovoltaic prediction is improved, the adaptability of a power grid to photovoltaic volatility is enhanced, and the energy scheduling efficiency and stability are optimized.

Description

Technical Field

[0001] The present invention relates to photovoltaic power prediction technology, and in particular to an ultra-short-term photovoltaic power prediction method, system, equipment and medium based on multivariate data fusion. Background Art

[0002] The WRF-Solar model is the first mesoscale professional meteorological model developed based on the WRF (Weather Research and Forecasting) modeling framework and is designed specifically for solar energy applications.

[0003] The highly volatile output of renewable energy sources such as wind and photovoltaic power generation, coupled with a mismatch between the temporal and spatial distribution of loads, significantly exacerbates the pressure on power grids to regulate frequency and peak loads. Against this backdrop, photovoltaic power forecasting has become a cornerstone technology for the safe, economical, and low-carbon operation of new power systems. Its core value lies in balancing the randomness of renewable energy with the controllability of the power grid, driving the transformation of the energy structure from a "source follows load" model to a "source-load interaction" model. As the proportion of photovoltaic installed capacity continues to increase, accurate photovoltaic forecasting technology is crucial.

[0004] Existing short-term photovoltaic forecasting technologies primarily rely on machine learning models (such as long short-term memory (LSTM), bidirectional long short-term memory networks (BLSTM), and temporal convolutional networks (TCN)). These models are combined with meteorological data correction (numerical weather prediction (NWP), convolutional neural networks (CNN), and grey relational analysis), multi-time-scale analysis, and data preprocessing techniques to perform photovoltaic power forecasting. However, this process involves significant uncertainty in meteorological data correction, resulting in insufficient accuracy in numerical weather forecasts. Localized meteorological fluctuations (such as short-term severe convection) are particularly difficult to capture, and absolute errors increase significantly in extreme weather conditions, thus impacting the accuracy of photovoltaic power forecasts. Summary of the Invention

[0005] In response to the deficiencies in the prior art, the present invention provides an ultra-short-term photovoltaic power prediction method, system, device and medium based on multivariate data fusion. By fusing multi-source heterogeneous data such as underlying surface data, real-time ground meteorological monitoring data, satellite monitoring data, and photovoltaic module operating status (such as temperature and attenuation rate), and combining the collaborative optimization of machine learning models and physical models, the present invention can effectively improve the accuracy of ultra-short-term numerical weather forecasts, reduce short-term prediction errors in extreme weather (such as short-term severe convection and aerosol mutations) and complex scenarios (such as cloudy and dusty weather), thereby improving the accuracy of ultra-short-term photovoltaic predictions, enhancing the grid's adaptability to photovoltaic fluctuations, and optimizing energy scheduling efficiency and stability.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] An ultra-short-term photovoltaic power prediction method based on multivariate data fusion includes the following steps:

[0008] Input underlying surface data, ground meteorological monitoring data, and satellite monitoring data into the WRF-Solar model to predict solar irradiance;

[0009] Solar irradiance, historical weather data, and historical data of photovoltaic power plants are input into the LSTM power prediction model to predict ultra-short-term photovoltaic power.

[0010] To optimize the above technical solutions, specific measures taken also include:

[0011] Furthermore, the WRF-Solar model is developed based on the WRF model, and the inputting of underlying surface data, ground meteorological monitoring data and satellite monitoring data into the WRF-Solar model specifically includes:

[0012] The coordinate system of the latest underlying surface data is converted to the geographic coordinate system required by the WRF model through the geographic information system. The latest underlying surface data is reclassified according to the WRF classification system, and the classified latest underlying surface data replaces the original underlying surface data in the WRF model.

[0013] Surface meteorological monitoring data and FNL global reanalysis data are integrated into the WRF model based on the four-dimensional variational assimilation method and spectral nudging method.

[0014] Based on the WRFDA system, the radiation data monitored by the Sunflower-8 satellite are integrated into the WRF model.

[0015] Furthermore, the underlying surface data is specifically a global land cover classification dataset, and the method for generating the global land cover classification dataset is:

[0016] Collect reflectance data from MODISTerra and Aqua satellites, classify and post-process the reflectance data to generate a global land cover classification dataset;

[0017] The classification method is supervised decision tree classification, maximum likelihood method, support vector machine or spectral angle mapping method, and the post-processing includes noise removal, classification result optimization, multi-temporal data fusion and quality control.

[0018] Furthermore, the four-dimensional variational assimilation method adopts an incremental form, and the cost function of the four-dimensional variational assimilation method is specifically:

[0019]

[0020] Where J(δx0) is the cost function of the four-dimensional variational assimilation method, δx0 is the analysis increment, and x0 b and x0 g They are the background field and the initial guess field at the initial moment, B0 represents the background error covariance matrix at the initial moment, and the observation field is divided into N sub-time windows within a time window, n is the sequence number of the sub-time window, d n is the new interest increment, H n and M n are the linearized observation operator and forward mode respectively, R0 represents the observation error covariance matrix at the initial moment;

[0021] The expression for the analytical increment is as follows:

[0022]

[0023] Where x0 is the analysis field;

[0024] The expression of innovation increment is as follows:

[0025]

[0026] Where y n 0 represents the observation field of the nth sub-time window, H n and M n are the nonlinear observation operator and forward mode respectively;

[0027]

[0028] Where H n and M n are the linearized observation operator and forward mode, x0 g is the initial guess field, δx0 is the analysis increment, H n and M n are the nonlinear observation operator and forward mode respectively, and x0 is the analysis field.

[0029] Furthermore, the SpectralNudging method is specifically as follows:

[0030] The FNL global reanalysis data is converted into a spectrum through fast Fourier analysis. Spectral analysis is performed, and the wave number is set. High-frequency waves greater than the set wave number are filtered out in space, while large-scale low-frequency waves are retained. Then, according to the set weights, the low-frequency waves are added to the model forecast field, thereby continuously bringing the simulation state closer to the large-scale driving state. This can be expressed as follows:

[0031]

[0032] Where α is the horizontal wind, potential temperature or geopotential height, and the value of α is obtained by interpolating the analysis values at adjacent moments, t is time, X(α) is the physical forcing term of the model, and G α is the Nudging coefficient, w(η) is the weight coefficient of Nudging in the vertical direction;

[0033] α 0,pq is the analytical field when the beams are p and q at the initial moment, α pq is the analysis field when the beams are p and q, i represents the time, x represents the background field, y represents the observation field, K pq Represents the approximation coefficient for different scales; K p , K q represent the wave vector components in the longitudinal and latitudinal directions, respectively; p and q represent the wave numbers in the longitudinal and latitudinal directions, respectively; P and Q are the Nudging wave numbers in the longitudinal and latitudinal directions, and their values are calculated based on the grid spacing, number of grid points, and wavelength. They are used to adjust the large-scale circulation field of the model to make it consistent with the real forcing field.

[0034] Furthermore, the LSTM power prediction model includes an input layer, an LSTM layer and an output layer. The LSTM layer is used to process the time dependency in the time series data. The LSTM layer has a two-layer structure and utilizes the hidden layer dimension to enhance the LSTM power prediction model's ability to remember and capture time series patterns.

[0035] The present invention also proposes an ultra-short-term photovoltaic power prediction system based on multivariate data fusion, comprising:

[0036] The WRF-Solar model is used to predict solar irradiance based on underlying surface data, ground-based meteorological monitoring data, and satellite monitoring data;

[0037] The LSTM power prediction model is used to predict ultra-short-term photovoltaic power based on solar irradiance, historical weather data, and historical data of photovoltaic power plants.

[0038] The present invention also proposes an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the ultra-short-term photovoltaic power prediction method based on multivariate data fusion as described above is implemented.

[0039] The present invention further provides a computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the ultra-short-term photovoltaic power prediction method based on multivariate data fusion as described above.

[0040] The beneficial effects of the present invention are as follows: based on the WRF-Solar model, the present invention integrates the underlying surface data of the study area and various real-time monitoring data (such as ground meteorological monitoring data, satellite monitoring data, etc.), and superimposes the use of machine learning methods on the basis of considering the microphysical changes of atmospheric clouds. It can capture local meteorological mutations (such as short-term severe convection), effectively improve the accuracy of short-term numerical weather forecasts, and thus improve the accuracy of ultra-short-term photovoltaic forecasts, and has good application prospects and promotion significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is the overall flow chart of the ultra-short-term photovoltaic power prediction method based on multivariate data fusion proposed in the present invention. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0043] Example 1

[0044] The present invention proposes an ultra-short-term photovoltaic power prediction method based on multivariate data fusion. The overall process of the method is as follows: Figure 1 As shown, the following steps are included:

[0045] Input underlying surface data, ground meteorological monitoring data, and satellite monitoring data into the WRF-Solar model to predict solar irradiance;

[0046] Solar irradiance, historical weather data, and historical data of photovoltaic power plants are input into the LSTM power prediction model to predict ultra-short-term photovoltaic power.

[0047] The WRF-Solar model is developed based on the WRF model. The underlying surface data, ground meteorological monitoring data, and satellite monitoring data are input into the WRF-Solar model as follows:

[0048] Underlying Data Fusion: The latest high-resolution underlying surface data (based on reflectance data from MODISTerra and Aqua satellites) was selected. A global land cover classification dataset, MODIS MCD12Q1, was generated through supervised classification (using a global prior knowledge base and field survey data to select spectrally representative training samples; subsequently, supervised decision tree classification was primarily employed, in combination with maximum likelihood, support vector machines, and spectral angle mapping). This was combined with post-processing techniques (including noise removal, classification result optimization, multi-temporal data fusion, and quality control). The coordinate system of the original high-resolution underlying surface data was converted to the geographic coordinate system required by the Mesoscale Numerical Weather Forecasting Model (WRF) using ArcGIS. The data was then reclassified according to the WRF classification system (USGS-24), replacing the original underlying surface data in the WRF model (the underlying surface data currently used in WRF is based on the underlying surface distribution from 20 years ago). By updating the underlying surface data, the accuracy of weather forecasts is improved.

[0049] Fusion of ground-based meteorological monitoring data: Based on the four-dimensional variational data assimilation (4DVar) and spectral nudging method (SN method), real-time meteorological monitoring data (temperature, humidity, air pressure, wind speed, wind direction, etc.) and FNL (Final Operational Global Analysis) global reanalysis data are fused into WRF. While correcting the large-scale bias of the WRF model, the meteorological data at the time of analysis is revised, further improving the accuracy of short-term weather forecasts.

[0050] Satellite monitoring data fusion: Based on the WRFDA (Weather Research and Forecasting model data assimilation system) system, the radiation data of the Sunflower-8 satellite is fused into WRF to correct the meteorological data at the time of analysis, further improving the accuracy of short-term weather forecasts.

[0051] MODIS: Moderate-resolution Imaging Spectroradiometer.

[0052] Terra: It is a morning orbit satellite in the U.S. Earth Observing System (EOS) program, mainly used for land and atmospheric observations.

[0053] Aqua: It is an afternoon orbit satellite in the EOS program, focusing on the study of ocean and atmospheric water cycles.

[0054] ArcGIS: ArcGeographic Information System, geographic information system.

[0055] The four-dimensional variational assimilation method adopts an incremental form. The cost function of the four-dimensional variational assimilation method is specifically:

[0056]

[0057] Where J(δx0) is the cost function of the four-dimensional variational assimilation method, δx0 is the analysis increment, and x0 b and x0 g They are the background field and the initial guess field at the initial moment, B0 represents the background error covariance matrix at the initial moment, and the observation field is divided into N sub-time windows within a time window, n is the sequence number of the sub-time window, d n is the new interest increment, H n and M n are the linearized observation operator and forward mode respectively, R0 represents the observation error covariance matrix at the initial moment;

[0058] The expression for the analytical increment is as follows:

[0059]

[0060] Where x0 is the analysis field;

[0061] The expression of innovation increment is as follows:

[0062]

[0063] Where y n 0 represents the observation field of the nth sub-time window, H n and M n are the nonlinear observation operator and forward mode respectively;

[0064]

[0065] Where H n and M n are the linearized observation operator and forward mode, x0 g is the initial guess field, δx0 is the analysis increment, H n and M n are the nonlinear observation operator and forward mode respectively, and x0 is the analysis field.

[0066] The incremental form of 4DVar linearizes the observation operator and the forward model, which not only reduces the amount of calculation but also improves the mathematical conditions for solving the cost function.

[0067] The SpectralNudging method is as follows:

[0068] The FNL global reanalysis data is converted into a spectrum through fast Fourier analysis. Spectral analysis is performed, and the wave number is set. High-frequency waves exceeding the set wave number are spatially filtered out, while retaining large-scale low-frequency waves. The low-frequency waves are then added to the model forecast field according to the set weights, thereby continuously moving the simulation state closer to the large-scale driving state, effectively reducing large-scale errors and preventing excessive high-frequency data from being over-corrected by the large-scale analysis field. Model variables can freely develop small and medium-scale processes while also reflecting smaller-scale characteristics. This can be expressed as follows:

[0069]

[0070] Where α is the horizontal wind, potential temperature or geopotential height, and the value of α is obtained by interpolating the analysis values at adjacent moments, t is time, X(α) is the physical forcing term of the model, and G α is the Nudging coefficient, w(η) is the weight coefficient of Nudging in the vertical direction;

[0071] α 0,pq is the analytical field when the beams are p and q at the initial moment, α pq is the analysis field when the beams are p and q, i represents the time, x represents the background field, y represents the observation field, K pq Represents the approximation coefficient for different scales; K p , K q represent the wave vector components in the longitudinal and latitudinal directions, respectively; p and q represent the wave numbers in the longitudinal and latitudinal directions, respectively; P and Q are the Nudging wave numbers in the longitudinal and latitudinal directions, and their values are calculated based on the grid spacing, number of grid points, and wavelength. They are used to adjust the large-scale circulation field of the model to make it consistent with the real forcing field.

[0072] The LSTM power prediction model is a deep learning model for processing time series data. Its core goal is to improve the accuracy of photovoltaic power forecasts by capturing long-term dependencies in time series data. The LSTM power prediction model consists of an input layer, an LSTM layer, and an output layer. The LSTM layer is used to process the temporal dependencies in time series data. The LSTM layer has a two-layer structure and utilizes a hidden layer dimension (hidden_size = 64) to enhance the LSTM power prediction model's ability to memorize and capture time series patterns.

[0073] Example 2

[0074] The present invention proposes an ultra-short-term photovoltaic power prediction system based on multivariate data fusion corresponding to the method of embodiment 1, comprising:

[0075] The WRF-Solar model is used to predict solar irradiance based on underlying surface data, ground-based meteorological monitoring data, and satellite monitoring data;

[0076] The LSTM power prediction model is used to predict ultra-short-term photovoltaic power based on solar irradiance, historical weather data, and historical data of photovoltaic power plants.

[0077] The implementation of each module and module function in the system is completely consistent with the steps of the method in Example 1, so it will not be repeated here.

[0078] Example 3

[0079] The present invention proposes an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the ultra-short-term photovoltaic power prediction method based on multivariate data fusion as described in Example 1 is implemented.

[0080] Example 4

[0081] The present invention provides a computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the ultra-short-term photovoltaic power prediction method based on multivariate data fusion as described in the first embodiment.

[0082] In the embodiments disclosed herein, computer storage media can be tangible media that can contain or store programs for use by or in conjunction with an instruction execution system, device, or apparatus. Computer storage media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or equipment, or any suitable combination of the foregoing. More specific examples of computer storage media can include electrical connections based on one or more lines, portable computer disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), optical fibers, portable compact disk read-only memories (CD-ROMs), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0083] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0084] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions based on the principles of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, various improvements and modifications that do not depart from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. An ultra-short-term photovoltaic power prediction method based on multivariate data fusion, characterized in that: The following steps are involved: Input underlying surface data, ground meteorological monitoring data, and satellite monitoring data into the WRF-Solar model to predict solar irradiance; Solar irradiance, historical weather data, and historical data of photovoltaic power plants are input into the LSTM power prediction model to predict ultra-short-term photovoltaic power.

2. The ultra-short-term photovoltaic power prediction method based on multivariate data fusion according to claim 1, characterized in that: The WRF-Solar model is developed based on the WRF model. Inputting the underlying surface data, ground meteorological monitoring data and satellite monitoring data into the WRF-Solar model specifically includes: The coordinate system of the latest underlying surface data is converted to the geographic coordinate system required by the WRF model through the geographic information system. The latest underlying surface data is reclassified according to the WRF classification system, and the classified latest underlying surface data replaces the original underlying surface data in the WRF model. Surface meteorological monitoring data and FNL global reanalysis data are integrated into the WRF model based on the four-dimensional variational assimilation method and spectral nudging method. Based on the WRFDA system, the radiation data monitored by the Sunflower-8 satellite are integrated into the WRF model.

3. The ultra-short-term photovoltaic power prediction method based on multivariate data fusion according to claim 2, characterized in that: The underlying surface data is specifically a global land cover classification dataset, and the method for generating the global land cover classification dataset is: Collect reflectance data from MODISTerra and Aqua satellites, classify and post-process the reflectance data to generate a global land cover classification dataset; The classification method is supervised decision tree classification, maximum likelihood method, support vector machine or spectral angle mapping method, and the post-processing includes noise removal, classification result optimization, multi-temporal data fusion and quality control.

4. The ultra-short-term photovoltaic power prediction method based on multivariate data fusion according to claim 2, characterized in that: The four-dimensional variational assimilation method adopts an incremental form, and the cost function of the four-dimensional variational assimilation method is specifically: Where J(δx0) is the cost function of the four-dimensional variational assimilation method, δx0 is the analysis increment, and x0 b and x0 g They are the background field and the initial guess field at the initial moment, B0 represents the background error covariance matrix at the initial moment, and the observation field is divided into N sub-time windows within a time window, n is the sequence number of the sub-time window, d n is the new interest increment, H n and M n are the linearized observation operator and forward mode respectively, R0 represents the observation error covariance matrix at the initial moment; The expression for the analytical increment is as follows: Where x0 is the analysis field; The expression of innovation increment is as follows: Where y n 0 represents the observation field of the nth sub-time window, H n and M n are the nonlinear observation operator and forward mode respectively; Where H n and M n are the linearized observation operator and forward mode, x0 g is the initial guess field, δx0 is the analysis increment, H n and M n are the nonlinear observation operator and forward mode respectively, and x0 is the analysis field.

5. The ultra-short-term photovoltaic power prediction method based on multivariate data fusion according to claim 2, characterized in that: The SpectralNudging method is specifically: The FNL global reanalysis data is converted into a spectrum through fast Fourier analysis. Spectral analysis is performed, and the wave number is set. High-frequency waves greater than the set wave number are filtered out in space, while large-scale low-frequency waves are retained. Then, according to the set weights, the low-frequency waves are added to the model forecast field, thereby continuously bringing the simulation state closer to the large-scale driving state. This can be expressed as follows: Where α is the horizontal wind, potential temperature or geopotential height, and the value of α is obtained by interpolating the analysis values at adjacent moments, t is time, X(α) is the physical forcing term of the model, and G α is the Nudging coefficient, w(η) is the weight coefficient of Nudging in the vertical direction; α 0,pq is the analytical field when the beams are p and q at the initial moment, α pq is the analysis field when the beams are p and q, i represents the time, x represents the background field, y represents the observation field, K pq Represents the approximation coefficient for different scales; K p , K q represent the wave vector components in the longitudinal and latitudinal directions, respectively; p and q represent the wave numbers in the longitudinal and latitudinal directions, respectively; P and Q are the Nudging wave numbers in the longitudinal and latitudinal directions, and their values are calculated based on the grid spacing, number of grid points, and wavelength. They are used to adjust the large-scale circulation field of the model to make it consistent with the real forcing field.

6. The ultra-short-term photovoltaic power prediction method based on multivariate data fusion according to claim 1, characterized in that: The LSTM power prediction model includes an input layer, an LSTM layer and an output layer. The LSTM layer is used to process the time dependency in time series data. The LSTM layer has a two-layer structure and utilizes the hidden layer dimension to enhance the LSTM power prediction model's ability to remember and capture time series patterns.

7. An ultra-short-term photovoltaic power prediction system based on multivariate data fusion, characterized in that: include: The WRF-Solar model is used to predict solar irradiance based on underlying surface data, ground-based meteorological monitoring data, and satellite monitoring data; The LSTM power prediction model is used to predict ultra-short-term photovoltaic power based on solar irradiance, historical weather data, and historical data of photovoltaic power plants.

8. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the ultra-short-term photovoltaic power prediction method based on multivariate data fusion as described in any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium storing a computer program, characterized in that: The computer program enables a computer to execute the ultra-short-term photovoltaic power prediction method based on multivariate data fusion according to any one of claims 1 to 6.

Images