A HYBRID APPROACH FOR SOLAR POWER PREDICTION BASED ON SATELLITE IMAGES AND SUPPORT VECTOR MACHINE

Abstract

The increasing penetration of solar energy into the main grid, driven by the rise of large-scale photovoltaic (PV) farms, has introduced challenges due to the variability of meteorological conditions, leading to fluctuations in power output. To address this, the paper proposes a solar power prediction model that combines satellite images with a Support Vector Machine (SVM) learning scheme. The model forecasts cloud motion vectors using satellite images of Atmospheric Motion Vectors (AMVs), enhancing the accuracy of solar power predictions. By analyzing four years of historical satellite data, a substantial dataset is configured for SVM training. The performance of this SVM-based model is compared to conventional time-series models and artificial neural networks (ANN) in terms of prediction accuracy. The results demonstrate that the proposed SVM-based approach outperforms traditional models, offering a more reliable method for predicting solar power output and mitigating the impact of weather-related fluctuations on the grid.

Specification

Description:FIELD OF INVENTION
The field of interest in developing a hybrid approach for solar power prediction that combines satellite imagery with Support Vector Machine (SVM) models. This method aims to enhance the accuracy of solar energy forecasts by integrating spatial data from satellite images with machine learning techniques, contributing to more efficient solar power management and renewable energy optimization.
BACKGROUND OF INVENTION
The growing reliance on solar energy as a renewable power source necessitates accurate prediction models to optimize energy production and grid management. Traditional methods for solar power prediction often rely on historical weather data and meteorological models, which may not fully capture the spatial variability and transient nature of solar irradiance. With advancements in remote sensing technology, satellite imagery provides high-resolution, real-time data on cloud cover, atmospheric conditions, and surface albedo, which are critical factors affecting solar energy generation. To address the limitations of existing models, this invention proposes a hybrid approach that integrates satellite images with Support Vector Machine (SVM) algorithms for solar power prediction. SVM, a robust machine learning technique, excels in handling nonlinear relationships and high-dimensional data, making it well-suited for modeling the complex interactions between atmospheric variables and solar irradiance. By incorporating satellite-derived features into the SVM framework, the proposed method enhances prediction accuracy, especially in regions with dynamic weather patterns. This hybrid approach is particularly beneficial for improving short-term solar power forecasts, enabling more efficient energy management and better integration of solar power into the electrical grid. It also offers potential applications in optimizing the operation of solar power plants and advancing the development of smart grid systems.
The patent application number 201941002461 discloses a energy efficient retrofitting kit design for converting combustible vehicle to solar powered.
The patent application number 02111043612 discloses a structure of a solar powered vehicle for mounting a plurality of solar modules.
The patent application number 202211025107 discloses an integrated, automated, solar power operated, floating fish-feeding assembly system.
The patent application number 202211032781 discloses a solar power integrated bidirectional non-isolated dc-dc converter for e-rickshaw.
SUMMARY
This invention presents a hybrid approach for solar power prediction that leverages the combined strengths of satellite imagery and Support Vector Machine (SVM) algorithms. The method is designed to improve the accuracy of solar energy forecasts by integrating real-time, high-resolution data from satellite images with the predictive capabilities of SVM, a machine learning model known for its ability to manage complex, nonlinear relationships. Satellite images provide critical information on factors such as cloud cover, atmospheric conditions, and surface reflectivity, all of which significantly influence solar irradiance and, consequently, solar power generation. Traditional forecasting models often overlook these spatial and temporal variations, leading to less reliable predictions. By extracting relevant features from satellite data and feeding them into the SVM, this hybrid approach captures the intricate interactions between environmental conditions and solar energy output more effectively than conventional methods. The invention is particularly useful for short-term solar power forecasting, crucial for grid operators and solar power plant managers who need precise data to optimize energy production, storage, and distribution. This approach not only enhances the reliability of solar power predictions but also supports the broader integration of renewable energy into the electrical grid, contributing to more sustainable and efficient energy management practices.
DETAILED DESCRIPTION OF INVENTION
The world is currently grappling with significant natural resource depletion and climate change, primarily due to the overuse of fossil fuels. To address these issues, there is a pressing need to shift towards alternative, renewable energy sources. The integration of renewable energy into the main power grid has steadily increased in recent years, and this trend is expected to accelerate significantly by 2030. Specifically, in South Korea, the proportion of solar energy among all energy sources is projected to rise gradually, reaching 14.1% by 2035.
Large-scale photovoltaic (PV) farms have been extensively deployed worldwide, with countries like Germany, China, and the U.S. leading the way. South Korea alone installed approximately 467 MW of PV plants in 2013. As the number of large-scale PV farms grows, the share of solar power in the overall power grid increases as well. However, the power output of these farms can be highly variable due to changing meteorological conditions, making accurate predictions of solar energy production essential for operators to effectively trade solar power in the energy market.
Solar power prediction technologies are thus crucial, as the accuracy of these predictions directly impacts the economic benefits derived from solar energy. Numerous studies have focused on developing prediction technologies for solar irradiance and PV power generation, which can be categorized into three main groups.
The first group involves technologies that predict solar irradiance or cloud index using satellite images. Since solar irradiance variability is largely influenced by cloud cover and movement-detected via satellite images-and to some extent by aerosols, these technologies are vital. Meteorological satellite images offer valuable insights into cloud motion and overall weather changes, with solar irradiance predictions utilizing cloud motion vectors (CMVs) as proposed in various studies.
The second group includes cloud detection technologies used to predict solar irradiance. Image processing techniques that analyze cloud movement and generate short-term forecasts of solar irradiance at ground level using total sky imagers (TSIs) have been suggested in several studies.
The third group focuses on solar power prediction technologies that employ machine learning methods. Time series models like ARMA and ARIMA have been applied to solar power prediction, along with spatial-temporal autoregressive models with exogenous input (ARX). However, these models face challenges in handling non-linear parameters such as cloud movement and meteorological variables. To overcome these limitations, artificial neural networks (ANNs) and support vector machines (SVMs) have been used to forecast global and horizon solar irradiance and PV system power generation.
Several studies have explored solar irradiance prediction using SVMs, mainly incorporating meteorological data like historical temperature and irradiance in the models. In contrast, this paper proposes a solar power prediction model based on various satellite images, including atmospheric motion vectors (AMVs), clouds, and irradiance, combined with an SVM learning approach. The key advantage of using satellite images in solar power prediction is the ability to observe cloud cover and movement on a macroscopic scale. Additionally, the SVM learning method enhances prediction accuracy by minimizing errors and improving the model's generalization capability. The proposed model first extracts AMVs from satellite images and then predicts cloud movement using these vectors. By analyzing four years of historical satellite images, a comprehensive dataset is created for SVM learning, enabling the model to predict future cloud cover and solar irradiance at multiple sites across South Korea within a range of 15 to 300 minutes (intra-day). The prediction accuracy of this model is compared with traditional time-series and ANN models, demonstrating its superior performance.
Meteorological satellite images for North-East Asia are provided by the Korea National Meteorological Satellite Center (NMSC). The NMSC launched Korea's first geostationary multi-purpose satellite, the Communication, Ocean, and Meteorological Satellite (COMS), on June 27, 2010. This satellite performs meteorological and ocean observations, along with communication services. Its primary missions include continuous monitoring of weather imagery, early detection of severe weather phenomena, and monitoring climate change and the atmospheric environment. The NMSC offers various types of satellite images, including raw, basic, and processed images, focusing on North-East Asia and the Korean Peninsula. Each image covering the Korean Peninsula consists of 1024 by 1024 pixels, with each pixel representing a ground resolution of approximately 1.7 kilometers. These images are typically updated every 15 minutes, though there is an approximate one-hour delay due to communication and processing procedures.
Three main types of satellite images are discussed in this context:
1. Atmospheric Motion Vector (AMV) Images:
AMV images provide comprehensive information about atmospheric motion, including wind direction and speed across lower, middle, and upper wind fields. These fields are represented by different colors-red, green, and blue-on the images. These images are crucial for analyzing wind patterns and require precise extraction of the atmospheric motion vectors through advanced image processing techniques.
2. Cloud Analysis Images:
The NMSC provides a range of cloud analysis images, which depict various cloud characteristics such as the amount, shape, and thickness of clouds. These images are essential for forecasting future cloud conditions and serve as important input variables for machine learning models.
3. Irradiance Images:
Irradiance images measure the amount of light reflected by the ground, expressed in watts per square meter (W/m²). These images are particularly useful for predicting photovoltaic (PV) power generation when combined with data on module temperature.
These satellite images are vital for various meteorological and environmental analyses, supporting accurate weather forecasting and environmental monitoring.

Figure 1: A standard model for wind direction and speed
Proposed AMV Extraction Method
This section outlines a method for extracting Atmospheric Motion Vectors (AMVs) to determine wind direction and speed. Before describing the method, we'll clarify some important terms and notations. Figure 2 illustrates an individual AMV vector, where the head, body, and tail are located at the top of the AMV wing (red circle), the bottom of the wing (green circle), and the end of the AMV line (blue circle), respectively. The midpoints between the head and body, and between the head and tail, are indicated by orange and purple diamonds, respectively.


Figure 2: A single AMV vector
AMV Extraction Process
The proposed method for extracting AMVs consists of four key steps:
1. Color Pixel Separation: The first step is to separate the red, green, and blue pixels from an AMV satellite image and store their coordinates in separate repositories for red, green, and blue AMVs. To demonstrate the process, we will focus on the extraction of blue AMVs in the following steps.
2. Individual AMV Extraction: In this step, individual blue AMVs are extracted from their repository. A pixel is randomly selected from the blue AMV repository, and the distances between this pixel and all other pixels in the repository are calculated. The closest pixel is added to a sub-repository for the single blue AMV, becoming the next reference point. This process continues under two constraints: (1) the distance between the current and next pixel must be below a specified threshold, and (2) the next pixel must be close to the central coordinate of the pixels already in the sub-repository. Once no more pixels satisfy these conditions, the extraction of that single blue AMV is complete, and its sub-repository is finalized. This process is repeated to extract all blue AMVs from the repository.
3. Head, Body, and Tail Identification: Next, we identify the head, body, and tail pixels for each blue AMV sub-repository. The two most distant pixels are identified as the head and tail, and their midpoint (purple diamond in Figure 2) is calculated. The pixel farthest from this midpoint, excluding the head and tail, is designated as the body of the AMV. The pixel closer to the body is designated as the head, while the other becomes the tail. Finally, a wind direction vector d=(dx, dy) is calculated from the body to the tail, along with the wind direction angle θ. This procedure is repeated for all sub-repositories.
4. Wind Speed Calculation: The final step is to calculate the wind speed for each blue AMV. First, the total number of pixels in a sub-repository is counted, and the midpoint (orange diamond in Figure 3) between the head and body is determined. The distances from this midpoint to all other pixels in the sub-repository are then calculated. If more than four pixels are within a certain distance from the midpoint, the initial wind speed is set to 25 m/s; otherwise, it is set to 5 m/s. The final wind speed is then adjusted based on the presence of speed bars (long bar: 5 m/s, short bar: 2 m/s).
This method provides a systematic approach for extracting AMVs and determining both wind direction and speed from satellite imagery.

Figure 3: The target and search areas on an AMV satellite image.
Extraction of Cloud Impact Factors from Search Areas
Target and Search Areas
To predict solar power for a specific region, we first define a target area and one or more surrounding search areas. The target area, where the solar power prediction is focused, is characterized by its coordinates (x target, y target), width w target, and height h target. The search areas, which are neighboring zones around the target area, are defined by their coordinates (x search,i, ysearch,i), width w search, and height h search. The accuracy and locality of the solar power prediction depend on the size of the target area and the characteristics of the surrounding search areas.
Cloud Factor Extraction
The next step involves extracting cloud factors from the search areas and determining how each search area influences the target area. Cloud factors are influenced by Atmospheric Motion Vectors (AMVs), which are divided into three color channels: Red (R), Green (G), and Blue (B). For each search area i and each color k, the wind direction, speed, and angle of the AMVs are denoted by di, k={dx(i,k), dy(i,k), and θ i,k, respectively.
Calculation of Cloud Impact Factor
It is assumed that the movement of clouds approximately follows the wind direction and speed, though the actual speed of clouds may differ from the wind speed. The optimal ratio of cloud speed to wind speed is denoted by α. This analysis assumes that new cloud formations or disappearances do not occur.
The overall AMV image includes the target area (indicated by a magenta box) and the search areas (indicated by cyan boxes). For each search area iii with color k, the average AMV parameters Ωi,k are calculated as follows:

where (xi,k(t),yi,k(t)) are the new center coordinates of the i-th search area with color k after t minutes, and ϵ represents the ground resolution of one pixel.
If the dimensions of the target and search areas are identical (w target = w search and h target = h search, the overlap between the target area and the moved search area is calculated as:

The cloud impact factor for the i-th search area on the target area after t minutes is calculated as:

The total cloud impact factor on the target area after t minutes is given by:

where I is the total number of search areas.
Training with Support Vector Machine Regression
We aim to train a solar power prediction model using SVM regression with historical satellite image data. Here's an overview of the process, including the parameter definitions and training procedures:
Parameters and Data Definition
1. Data Set Definition:

2. Objective:
We aim to model the complex, nonlinear relationship yi=g(xi) by approximating it with a function f(xi) using SVM regression.
Support Vector Machine (SVM) Regression
1. SVM Overview:
o Original Concept: Developed by Vapnik, SVM is known for its high performance in classification and regression tasks. It not only minimizes error but also maximizes the margin between classes.
o Regression Function: In SVM regression, the function is formulated as f(xi)=wTxi+b, where www and xi are vectors in RK, and b is a scalar.
o Optimization Problem: The goal is to find the optimal w by solving:

Kernel Method: For nonlinear regression, we use the kernel trick to map input vectors to a higher-dimensional space. The function becomes:

Training Methods
1. Cloud Amount Prediction:
o Input Vector: , where C represents current cloud amounts, and Γ(t) is the cloud impact factor for the prediction horizon t. The output yC is the amount of clouds at the target area after t minutes.
o Clear-Sky Index Calculation: The clear-sky index η is computed as:

Irradiance Prediction:
• Extended Input Vector: For predicting irradiance, include additional data: xΨ = [C, Γ(t), ω, Ψtarget(0)]T, where ω is the solar altitude angle, and Ψtarget(0) is the current irradiance.
• Output Value: The target irradiance yΨ is predicted based on:

Performance Evaluation
We analyze the performance of the proposed Support Vector Machine (SVM)-based prediction models for cloud coverage and irradiance. We start by examining the statistical characteristics of cloud coverage and irradiance data in South Korea, followed by evaluating the accuracy of the prediction models using metrics such as Root Mean Square Error (RMSE), Mean Relative Error (MRE), and the coefficient of determination (R²). These metrics are defined as follows:
• RMSE:

• R²:


We use four years of satellite data from the National Meteorological Satellite Center (NMSC) to configure extensive datasets for machine learning (ML) training and testing.
DETAILED DESCRIPTION OF DIAGRAM
Figure 1: A standard model for wind direction and speed
Figure 2: A single AMV vector
Figure 3: The target and search areas on an AMV satellite image. , Claims:1. A Hybrid Approach for Solar power prediction based on satellite images and support vector machine claims that the proposed model utilizes a large dataset of historical satellite images, spanning four years, combined with an SVM (Support Vector Machine) learning scheme for predicting solar power.
2. The model processes extensive historical satellite imagery to generate numerous input and output data sets for training the SVM algorithm.
3. It can forecast both the future cloud cover and solar irradiance across multiple locations in South Korea, with prediction intervals ranging from 15 to 300 minutes (intraday).
4. The SVM-based prediction model achieves higher accuracy compared to traditional time-series forecasting models and artificial neural network (ANN) models.
5. The accurate predictions provided by the model are beneficial for grid operations, particularly in load following, which helps in balancing supply and demand.
6. The prediction data from the model can be integrated into Energy Management Systems (EMS) within smart grids for more efficient energy management.
7. The approach involves a comprehensive analysis of historical satellite images, which enriches the data used for training and improves prediction reliability.
8. The use of SVM learning enhances the model's ability to generalize and make accurate predictions based on the input data.
9. The model is specifically designed for intraday solar power forecasting, addressing short-term variations in solar irradiance and cloud cover.
10. The proposed SVM-based model demonstrates superior performance in predicting solar power compared to other existing forecasting methods, establishing it as a reliable tool for solar energy management.

Documents