Doctoral School of Electrical Engineering 

BME-VIK, Department of Electric Power Engineering, HUN-REN Center for Energy Research, Environmental Physics Department

Supervisor: Dr. GROMA Veronika, Dr. HARTMANN Bálint

Solar power production estimation combining artificial intelligence and traditional image processing

Introducing the research area

Thanks to the rapid expansion of solar energy production over the past decade, the use of various solar irradiation forecasts has become more significant. Solar power generation is determined by highly variable solar radiation, as rapid and irregular cloud movement complicates accurate forecasting, presenting a considerable challenge for grid operators. Our research aims to develop a predictive methodology that, based on images captured of the sky, forecasts expected solar radiation for short time intervals. In our study, we combine artificial intelligence and traditional image processing algorithms, as these two methods complement each other, compensating for each other's drawbacks.

 

 

Figure 1. All-sky camera at the Centre of Energy Research 

Brief introduction of the research place

HUN-REN Center for Energy Research, Environmental Physics Laboratory, Energy Strategy and Environmental Impacts Research Group

Our research group consists of experts with diverse professional backgrounds (electrical engineers, energy engineers, architects, meteorologists, physicists, and geographers) tailored to the diversity of interdisciplinary energy science's fundamental and applied research questions. We work across five research focus areas, including energy system security, electrical energy, energy meteorology and climatology, energy statistics, as well as building and urban energy.

History and context of the research

In line with national and international decarbonization goals, the domestic solar capacity has significantly expanded in recent years, reaching 5000 MWp last year [1]. The increase in the number of solar panels has a radical impact on the electricity grid; it risks grid stability, burdens the grid infrastructure, and causes fluctuating or excessively high voltage for consumers. These effects are occurring more and more frequently, highlighting the importance of ultra-short-term (maximum one-hour time horizon) production estimation in addressing these issues and understanding the network impact of solar panels.

 

Solar production is fundamentally determined by cloud cover, so short-term production estimation methods rely on tracking and forecasting cloud cover, typically achieved through analysis of sky camera images. In the Environmental Physics Laboratory of the Energy Science Research Center, we have a continuously operating sky camera (Figure 1), which captures images of the sky as a percentage (Figure 2). These images are analyzed to generate solar radiation forecasts, supplemented by additional measurements, such as diverse meteorological data recorded by weather stations located near the camera (e.g., wind speed, humidity, temperature).

The research goals, open questions

 

Figure 2 Image of the all-sky camera

 

The research aims to develop a methodology for forecasting solar radiation, the parameter directly influencing the production of solar panels. To achieve accurate forecasts, we need to integrate cloud cover and meteorological measurements as input, as these two datasets are generally insufficient for precise predictions individually [2]. Since there is a complex relationship between cloud cover, meteorological measurements, and solar radiation that cannot be described by simple relationships, we employ a data-driven approach instead of model-based techniques. Our goal is to develop a neural network model capable of efficiently processing large amounts of weather data and images and recognizing complex patterns between the input data and solar radiation (Figure 3).

 

Figure 3 The evolution of meteorological parameters and sky characteristics on four different weather days. The images, in sequence, depict: a sunny day with a clear sky, a day with varying cloud cover, a cloudy day, and a heavily overcast winter day. The images illustrate that the two components of solar radiation (global radiation and diffuse radiation) are primarily determined by the cloudiness of the sky. However, factors such as the proportion of cloud inhomogeneity (which refers to the ratio of thick, dark clouds within the total cloud cover) also significantly influence solar radiation, even under similar cloud cover conditions.

Methods

One of the most popular methods for radiation estimation is the application of artificial neural networks. The advantage of neural networks is that they enable the combination and simultaneous processing of various measurement data (such as images, weather data, and timestamps) and can also consider temporal changes, making them suitable methods for radiation prediction. However, a significant obstacle to the application of neural networks is the high computational demand. This is because the number of parameters in a network trained for a complex scientific task is on the order of millions, making both training and deployment costly, requiring computing infrastructure equipped with graphical processors.

 

To address this issue, we have developed a methodology that first analyzes sky camera images using low-computational algorithms, quantifying the main characteristics of the sky that determine solar radiation. Such characteristics include the percentage of cloud cover relative to the total sky, the transparency of clouds, or the fragmentation of cloud cover. For extracting these features, we apply simple algorithms commonly used in classical image processing, such as thresholding, searching for connected regions, or color analysis. This processing is highly efficient, resulting in obtaining the basic cloud characteristics for each image.

 

This method can be considered a compression technique, producing extracted data representing the sky from sky camera images, significantly reducing the amount of data. The pre-processed data is then provided as input to the neural network, requiring it to process much less information compared to directly providing sky images for processing. The reduced amount of data also determines the complexity and number of parameters of the network. As a result, the neural networks we use can be trained in a short time even on a personal computer. The schematic structure of the network is shown in Figure 4.

 Figure 4 The schematic structure of the developed neural network is as follows. At the input, we provide data from various sources, and at the output, we obtain two components of solar radiation: global radiation and diffuse radiation. The network is trained using actual radiation data measured by weather stations. For training and evaluation, we utilized data from two years, totaling 300,000 data points.

Results

Based on the data measured by weather stations, we compiled weather scenarios to understand how individual meteorological data influence the accuracy of our predictions. Our goal was to identify the meteorological measurements essential for accurate forecasting and assess which measurements could be omitted without affecting the results. We used different combinations of data in various scenarios, considering, for instance, one-minute average values and standard deviations for the same period in the first scenario, while using less data in other scenarios. During training, we utilized data points from two full years, totaling over 300,000 one-minute samples.

 

The results indicate that while the composition of meteorological data has no significant impact on accuracy, the inclusion of sky characteristics significantly improves accuracy compared to using only meteorological data (Figure 5). With the use of sky characteristics, the network better tracks sudden drops in solar radiation due to cloud effects.

 

We compared neural networks trained on different scenarios in terms of accuracy and resource requirements with more complex Vision Transformer-based neural networks developed by the research group, which utilize full camera images [3]. Compared to the network that processes full images, the accuracy of the prediction does not deteriorate; however, there is a significant difference in the required resources and the number of network parameters. Training the Vision Transformer network on the same personal computer takes several days, whereas the network using sky characteristics completes in about two hours. One advantage of this cheaper and more energy-efficient method is the ease of its integration into resource-constrained systems, which aspect is becoming increasingly important with the growing popularity of edge computing applications [4].

 

Expected impact and further research

Based on the short-term forecast of solar radiation and solar panel production, there are various ways to intervene in the operation of the electricity grid. The intervention aims to improve the efficiency of energy delivery and the quality of service. With accurate forecasts, sudden drops in production due to cloud cover can be avoided through the control of spinning reserves or energy storage systems, and in cases of overproduction, solar panels can be regulated as needed [5].

 

In the current phase of the research, our goal was to achieve the most accurate estimation of solar radiation. However, our plans include developing a predictive model for multiple time intervals using feedback neural networks. To gain a fuller understanding of the model's performance, we intend to conduct experiments under different weather conditions.

Figure 6  The evolution of the components of solar radiation (global radiation and diffuse radiation) on different weather days using sky features and without sky features. Tracking sudden radiation drops on days with variable cloudiness poses a challenge, yet in most cases, the network accurately estimates radiation. Sky features enable much more precise estimation as they carry information about the cloud cover.

 

Publications, references, links

List of corresponding own publications.

[1] Barancsuk, L., Groma, V., Günter, D., Osán, J., Hartmann, B., Estimation of Solar Irradiance Using a Neural Network Based on the Combination of Sky Camera Images and Meteorological Data. Energies, Energies, Vol. 17(2), pp. 438., (2024)

 

[2] Barancsuk, L.; Günter, D., Groma V., Hartmann B., Fotovillamos termelésbecslés támogatása multimodális neurális háló segítségével égboltkameraképek és különböző konfigurációjú időjárási adatok alapján, XIII. Mechwart András Ifjúsági Találkozó: Conference Proceedings, Siófok, Hungary: Magyar Elektrotechnikai Egyesület, Mechwart András Ifjúsági Társaság 181 p. pp. 7–19., 13 p. (2023)

 

[3] Günter, D., Barancsuk, L., Groma, V., Sinkovics, B., Az égboltkamerák szerepe a fotovillamos termelés rövidtávú előrejelzésében., Elektrotechnika, Vol. 116, pp. 7–9, (2023)

 

[4] Barancsuk, L.; Günter, D., Groma V., Fotovillamos termelésbecslés deep learning alapon, hagyományos képfeldolgozási módszerek integrálásával., Elektrotechnika, accepted for publication, (2024)

 

Table of links.

List of references.

[1] MAVIR, “Naptermelés becslés és tény adatok” Link: https://mavir.hu/web/mavir/naptermeles-becsles-es-teny-adatok, 2023, Time of access: 2024. 03. 30.

 

[2] F. Barbieri, S. Rajakaruna, A. Ghosh, “Very short-term photovoltaic power forecasting with cloud modeling: A review” Renewable and Sustainable Energy Reviews, vol. 75, pp. 242–263, 2017

 

[3] Barancsuk, L.; Günter, D., Groma V., Hartmann B., Fotovillamos termelésbecslés támogatása multimodális neurális háló segítségével égboltkameraképek és különböző konfigurációjú időjárási adatok alapján, XIII. Mechwart András Ifjúsági Találkozó: Konferenciakiadvány Siófok, Hungary: Magyar Elektrotechnikai Egyesület, Mechwart András Ifjúsági Társaság 181 p. pp. 7–19., 13 p. (2023)

 

[4] E.A. Papatheofanous, V. Kalekis; G. Venitourakis, F. Tziolos; D: Reisis,  “Deep Learning-Based Image Regression for Short-Term Solar Irradiance Forecasting on the Edge” Electronics, vol. 11, no. 3794, 2022

 

[5] R. Samu, M. Calais, G. Shafiullah, M. Moghbel, M. A. Shoeb, B. Nouri, and N. Blum, “Applications for solar irradiance nowcasting in the control of microgrids: A review” Renewable and Sustainable Energy Reviews, vol. 147, p. 111187, 2021.