Logo: to the web site of Uppsala University

Solar photovoltaics (PV) is currently the fastest growing type of electrical energy generation. A substantial share is distributed, and key information - such as their installed capacity, precise location, tilt, and azimuth - are often lacking or inaccurate. Therefore, obtaining accurate data on existing PV systems become increasingly critical to determine optimal locations for adding new PV capacity, in terms of ensuring grid stability. Recent advances in identifying and segmenting solar energy systems, using aerial imagery, point to a logical next step; enhancing the modelling of tilts and azimuths, as these influence the power output significantly. Therefore, a method is proposed that derives the tilt and azimuth of solar energy systems using Light Detection and Ranging (LiDAR) data. Polygons representing solar energy systems, identified in aerial images, are orthorectified to LiDAR data and then linear regression is applied to determine the orientation. The method is evaluated for 3'500 Swedish solar energy systems previously identified in aerial images, with a manually derived ground truth azimuth dataset. For 91%-95% of the systems, the model accurately estimated the azimuth within a margin of 3 degrees. Furthermore, the distribution of azimuths was more narrow for solar thermal systems than for PV systems. Aground truth of the tilt fora subset of 39 systems gave a mean absolute error of 3.6 degrees. The proposed method is believed to provide more accurate PV metadata to, e.g., aggregators and grid operators, enabling more precise PV power simulations and forecasts, in turn leading to better grid operation and planning.

A significant challenge is to determine the specific services Battery Energy Storage System (BESS) should provide to maximize profits. This study investigates the most profitable markets and sizes of BESS with utility-scale solar Photovoltaics (PV) power plants using techno-economic analysis frameworks. The objective is to maximize profitability in energy and frequency markets, focusing on primary regulation and day-ahead markets for Sweden and Germany. The inputs are historical market prices and frequency data, as well as real measurement PV power data. The results show that adding a BESS to an existing PV park does not result in a lower payback period than if implementing a stand-alone BESS. However, the payback period differs between Sweden and Germany during 2023, i.e., being 1.8 and 6.8 years, respectively. This is explained by the lower frequency market prices for Germany compared to Sweden. The technical results indicate that the BESS energy capacity after 10 years of operation is approximately 83% for Germany, whereas, for Sweden, it is around 87%. Also, combining the operating of BESS on primary regulation and day-ahead markets showed a 6-year payback period with a slight increase in loss of energy capacity (from 83 to 80%) for Germany. Moreover, combining various PV-BESS sizes showed a discrepancy in economic and technical metrics for the BESS in Germany, resulting in a best-case of a 6-year payback period. A sensitivity analysis, which examines a drop in the frequency control prices in the future relative to 2023 (by 20% and 50% for Germany and Sweden, respectively), reveals an increase in the payback period for both countries by approximately 1 year.

This paper presents a first step in the field of probabilistic forecasting of co-located wind and photovoltaic (PV) parks. The effect of aggregation is analyzed with respect to forecast accuracy and value at a co-located park in Sweden using roughly three years of data. We use a fixed modelling framework where we post-process numerical weather predictions to calibrated probabilistic production forecasts, which is a prerequisite when placing optimal bids in the day-ahead market. The results show that aggregation improves forecast accuracy in terms of continuous ranked probability score, interval score and quantile score when compared to wind or PV power forecasts alone. The optimal aggregation ratio is found to be 50%–60% wind power and the remainder PV power. This is explained by the aggregated time series being smoother, which improves the calibration and produces sharper predictive distributions, especially during periods of high variability in both resources, i.e., most prominently in the summer, spring and fall. Furthermore, the daily variability of wind and PV power generation was found to be anti-correlated which proved to be beneficial when forecasting the aggregated time series. Finally, we show that probabilistic forecasts of co-located production improve trading in the day-ahead market, where the more accurate and sharper forecasts reduce balancing costs. In conclusion, the study indicates that co-locating wind and PV power parks can improve probabilistic forecasts which, furthermore, carry over to electricity market trading. The results from the study should be generally applicable to other co-located parks in similar climates.

An important aspect in designing co-located wind and solar photovoltaic hybrid power plants is the sizing of the energy converters to achieve as efficient power smoothening as possible. In this study, the ratio of wind- and photovoltaic energy converters in a hybrid power plant is determined by minimizing the overall stored energy that is needed to facilitate constant power output. Using Fourier transform the variability is isolated at predefined time scales that are relevant for grid integration. For the investigated time scales, energy and power ratings for energy storages are determined to counteract the variability. The resulting configuration is the one that is able to achieve constant power output with minimum stored energy. It is shown that co-locating wind- and photovoltaic energy converters smoothen seasonal energy generation, and reduce the energy storage need in both the diurnal and seasonal time scales. A case study for south-eastern Sweden is presented where the wind- \& solar hybrid plant configuration that minimizes the energy storage need and therefore most closely resembles constant output power is determined. It is found that a ratio of approximately 40-45\% wind power in the hybrid power plant yields the lowest need for energy storage. The presented method is valid for any number of co-located energy sources, and can also be extended to sizing of hybrid power systems.

The electrical system is currently undergoing a transition, where newhigh-power flexible loads, e.g., electric vehicles (EVs), are penetrating residential areas, and distributed power production from e.g., photovoltaic (PV) panels are also rapidly increasing in the low voltage (LV) grid. This transition requires new modeling methods to accurately predict the vulnerabilities and the needs to upgrade the current grid. A methodology to utilize spatiotemporal Markov based models of PV and EV charging to evaluate the impacts of these technologies on the high voltage (HV)/medium voltage (MV) substations is presented in this report. Furthermore, a case study on a large Swedish city was made. In this case study, penetrations of 100% for the EVs and PV were simulated. The results indicated that EV charging increases the peak load in the city by up to 18%–28%, and the peak load in the substations increased by up to 55%. During July, the PV yield was at most 45% of the winter consumption peak in the city and the summer-time net in-feed was at most 77% in any of the primary substations. Only 3 out of 10 substations experienced overloading events, and in all but one substation these events were shorter than 17 h/year. These overloading have negligible impact on the life-time of the main transformers as they predominantly occur when the ambient temperature is low. To avoid expensive upgrades to the MV transformers, the reserve transformers in the substations can be used to alleviate these overloading incidences. This solution however will not solve hosting capacity limitations in the underlying grid.

Statistics on installed solar energy systems (SES) play a crucial role in the solar energy industry, providing valuable information for a wide range of stakeholders, such as policy makers, authorities, and financial evaluators. For example, grid operators rely on accurate data on photovoltaic penetration levels to ensure the quality and stability of the power supply. In this research, we present an automatic approach helping generate these statistics using deep learning and image processing techniques. Our proposed model is a machine learning approach that utilizes a specific architecture of convolutional neural networks (CNN) called the "U-net'' to detect SES from aerial images. We experimented different network settings to enhance the SES identification performance.In this study, the model was evaluated using two datasets from different locations, one from Sweden and one from Germany. Additionally, the model was trained and tested on a combination of both datasets. The impact of image resolution was also examined. The experimental results show that this architecture performs better than many recent CNN models that have been proposed in the literature for the task of SES identification from aerial images. To make it easy for others to replicate our findings, We have shared all the scripts, software, and dependencies required for running the model in this paper, along with instructions on how to use it in Appendix A.

As a mean to monitor the rapid expansion of the highly decentralized PV market, identifying solar energy systems in aerial imagery by deep machine learning, is a research field that is getting increasing interest. One general challenge in the field is to create testing data of high quality that are representative of the end-use application. In this study we use the open source convolutional neural network developed within the DeepSolar project and apply it in the country of Sweden, for the purpose of generating market statistics, by scanning three complete municipalities for small decentralized photovoltaic and solar thermal systems. The evaluation of the performance is done against a highly accurate ground truth, which was created by cross-checking the classification results with the inventory of the local distribution system operators and the database of photovoltaic systems that have received a capital subsidy in Sweden, and combining that with physical onsite inspections. A process of generate additional training data and re-training the algorithm after each municipality scan was developed, which successively improved the accuracy, resulting in that 95% of all detectable photovoltaic, excluding building integrated and vertical systems, and 80% of all detectable solar thermal systems were correctly identified in the last municipality scan. The accurate ground truth allowed a quantification of why some systems are not detected. The generated dataset of solar energy systems could be connected to existing building and property inventories, which allowed creation of market segment statistics with remarkably high detail information.

One of the essential epistemic uncertainties that has not yet been studied enough for distributed photovoltaic systems is the azimuth and tilt of rooftop photovoltaic panels, as previous studies of grid impacts and hosting capacity have tended to assume uniform and optimal roof facet conditions. In this study, rooftop facet orientation distributions are presented and analyzed for all single-family buildings in the Swedish city of Uppsala, based on LiDAR-based data that consist of every roof facet from the around 13,500 single-family buildings in the city. From these distributions, novel methods to proportionally include less suitable roofs for every penetration level are proposed using a simple method based on normal and uniform probability density functions, and are tested for both time-series and stochastic hosting capacity analysis. The results show that under the assumption that the best roof facets are utilized first, a uniform distribution for rooftop facet azimuth and a normal distribution for rooftop facet tilt with parameters that depend linearly on the penetration level were shown to be accurate. The hosting capacity simulations demonstrate how the proposed methods perform significantly better in estimating the photovoltaic hosting capacity than the more common simplified methods for both time-series and stochastic hosting capacity analysis. The proposed model could help distribution system operators as well as researchers in this area to model the rooftop facet orientation uncertainty better and improve the quality of aggregated photovoltaic generation models and hosting capacity analyses.