Rozprawy doktorskie na temat „SOLAR POWER FORECASTING”

Solar energy is a promising environmentally-friendly energy source. Yet its variability affects negatively the large-scale integration into the electricity grid and therefore accurate forecasting of the power generated by PV systems is needed. The objective of this thesis is to explore the possibility of using machine learning methods to accurately predict solar power. We first explored the potential of instance-based methods and proposed two new methods: the data source weighted nearest neighbour (DWkNN) and the extended Pattern Sequence Forecasting (PSF) algorithms. DWkNN uses multiple data sources and considers their importance by learning the best weights based on previous data. PSF1 and PSF2 extended the standard PSF algorithm deal with data from multiple related time series. Then, we proposed two clustering-based methods for PV power prediction: direct and pair patterns. We used clustering to partition the days into groups with similar weather characteristics and then created a separate PV power prediction model for each group. The direct clustering groups the days based on their weather profiles, while the pair patterns consider the weather type transition between two consecutive days. We also investigated ensemble methods and proposed static and dynamic ensembles of neural networks. We first proposed three strategies for creating static ensembles based on random example and feature sampling, as well as four strategies for creating dynamic ensembles by adaptively updating the weights of the ensemble members based on past performance. We then explored the use of meta-learning to further improve the performance of the dynamic ensembles. The methods proposed in this thesis can be used by PV plant and electricity market operators for decision making, improving the utilisation of the generated PV power, planning maintenance and also facilitating the large-scale integration of PV power in the electricity grid.

The increased competitiveness of solar PV panels as a renewable energy source has increased the number of PV panel installations in recent years. In the meantime, higher availability of data and computational power have enabled machine learning algorithms to perform improved predictions. As the need to predict solar PV energy output is essential for many actors in the energy industry, machine learning and time series models can be employed towards this end. In this study, a comparison of different machine learning techniques and time series models is performed across five different sites in Sweden. We find that employing time series models is a complicated procedure due to the non-stationary energy time series. In contrast, machine learning techniques were more straightforward to implement. In particular, we find that the Artificial Neural Networks and Gradient Boosting Regression Trees perform best on average across all sites.
Sänkta produktionskostnader och ökad effektivitet har de senaste åren gjort solceller till ett attraktivt alternativ som energikälla. Detta har lett till en stor ökning av dess användning runt om i världen. Parallellt med denna utveckling har större tillgänglighet av data samt datorers förbättrade beräkningskapacitet möjliggjort förbättrade prediktionsresultat för maskininlärningsmetoder. Det finns för många aktörer anledning att intressera sig för prediktion av solcellers energiproduktion och från denna utgångspunkt kan maskininlärningsmetoder samt tidsserieanalys användas. I denna studie jämför vi hur metoder från de båda fälten presterar på fem olika geografiska områden i Sverige. Vi finner att tidsseriemodeller är komplicerade att implementera på grund av solcellernas icke-stationära tidsserier. I kontrast till detta visar sig maskininlärningstekniker enklare att implementera. Specifikt finner vi att artificiella neurala nätverk och så kallade Gradient Boosting Regression Trees presterar bäst i genomsnitt över de olika geografiska områdena.

The purpose of this bachelor thesis is to investigate different outcomes of the usage of photovoltaic (PV) power for electric vehicle (EV) charging adjacent to workplaces. In the investigated case, EV charging stations are assumed to be connected to photovoltaic systems as well as the electricity grid. The model used to simulate different scenarios is based on a goal of achieving constant power exchange with the grid by adjusting EV charging to a solar irradiance forecast. The model is implemented in MATLAB. This enables multiple simulations for varying input parameters. Data on solar irradiance are used to simulate the expected PV power generation. Data on driving distances are used to simulate hourly electricity demands of the EVs at the charging stations. A sensitivity analysis, based on PV irradiance that deviates from the forecast, is carried out. The results show what power the grid needs to have installed capacity for if no PV power system is installed. Furthermore, appropriate PV power installation sizes are suggested. The suggestions depend on whether the aim is to achieve 100 percent self-consumption of PV generated power or full PV power coverage of charging demands. For different scenarios, PV power installations appropriate for reducing peak powers on the grid are suggested. The sensitivity analysis highlights deviations caused by interference in solar irradiance.

The optimization of photovoltaic (PV) power generation system requires an accurate system performance model capable of validating the PV system optimization design. Currently, many commercial PV system modeling programs are available, but those programs are not able to model PV systems on a distorted ground level. Furthermore, they were not designed to optimize PV systems that are already installed. To solve these types of problems, this thesis proposes an optimization method using model simulations and a MATLAB-based PV system performance model. The optimization method is particularly designed to address partial shading issues often encountered in PV system installed on distorted ground. The MATLAB-based model was validated using the data collected from the Cal Poly Gold Tree Solar Field. It was able to predict the system performance with 96.4 to 99.6 percent accuracy. The optimization method utilizes the backtracking algorithm already installed in the system and the pitch distance to control the angle of the tracker and reduces solar panels partial shading on the adjacent row to improve system output. With pitch distances reduced in the backtracking algorithm between 2.5 meters and 3 meters, the inverter with inter-row shading can expect a 10.4 percent to 28.9 percent increase in power production. The implementation and calibration of this optimization method in the field this spring was delayed due to COVID-19. The field implementation is now expected to start this summer.

The electrical grid can no longer be considered a unidirectional means of distributing energy from conventional plants to the final users, but a Smart Grid, where strong interaction between producers and users takes place. In this context, the importance of independent renewable generation is constantly increasing, and new tools are needed in order to reliably manage conventional power plant operation, grid balancing, real-time unit dispatching, demand constraints and energy market requirements. This dissertation is focused on two aspects of this general problem: cost-optimal management of smart buildings in a Demand-Response framework, and estimation of photovoltaic generation forecasting models. In the first part of this thesis a novel Model Predictive Control approach for integrated management of HVAC, electrical and thermal storage, and photovoltaic generation in building is presented. The proposed methodology also considers participation of the building in a Demand-Response program, which allows the consumer to become an active player in the electricity system. The related optimization problems turn out to be computationally appealing, even uncertainty sources is also addressed by means of a two-step procedure. The second part deals with the problem of estimating photovoltaic generation forecasting models in scenarios where measurements of meteorological variables (i.e., solar irradiance and temperature) at the plant site are not available. This scenario is relevant to electricity network operation, when a large number of photovoltaic plants are deployed in the grid. In particular, two methods have been developed. The first approach makes use of raw cloud cover data provided by a weather service combined with power generation measurements to estimate the parameters of a novel class of models. The second approach is based on a set of tests performed on the generated power time series aimed at detecting data portions that were generated under clear sky conditions. These data are then used for fit the parameters of the PVUSA model to the theoretical clear sky irradiance. All the methods covered in this thesis have been extensively validated either using industry-standard simulation frameworks or via experiments performed on real data.

As the ‘green’ energy movement continues to gain momentum, photovoltaic generation is becoming an increasingly popular source for new power generation. The primary focus of this paper is to demonstrate the benefits of close-to real-time cloud sensing for Photovoltaic generation. In order to benefit from this close-to real-time data, a source of cloud cover information is necessary. This paper looks into the potential of point insolation sensors to determine overhead cloud coverage. A look into design considerations and economic challenges of implementing such a monitoring system is included. The benefits of cloud location sensing are examined using computer simulations to target important time-scales and options available to plant operators. Finally, the economics of advanced forecasting options will be examined in order to determine the benefit to plant operators.

The increasing penetration of renewable energy sources into the electricity generating mix poses challenges to the operational performance of the power system. Similarly, the push for energy efficiency and demand response—i.e., when electricity consumers are encouraged to alter their demand depending by means of a price signal—introduces variability on the consumption side as well. Forecasting is generally viewed as a cost-efficient method to mitigate the adverse effects of the aforementioned energy transition because it enables a grid operator to reduce the operational risk by, e.g., unit-commitment or curtailment. However, deterministic—or point—forecasting is currently still the norm. This thesis focuses on probabilistic forecasting, a method with which the uncertainty ac- companying the forecast is expressed by means of a probability distribution. In this framework, the thesis contributes to the current state-of-the-art by investigating properties of probabilistic forecasts of PV power production, electricity consumption and net load at the residential and distribution level of the electricity grid. The thesis starts with an introduction to probabilistic forecasting in general and two models in specific: Gaussian processes and quantile regression. The former model has been used to produce probabilistic forecasts of PV power production, electricity consumption and net load of individual residential buildings—particularly challenging due to the stochasticity involved— but important for home energy management systems and potential peer-to-peer energy trading. Furthermore, both models have been utilized to investigate what effects spatial aggregation and increasing penetration have on the predictive distribution. The results indicated that only 20- 25 customers—out of a data set containing 300 customers—need to be aggregated in order to improve the reliability of the probabilistic forecasts. Finally, this thesis explores the potential of Gaussian process ensembles, which is an effective way to improve the accuracy of the forecasts.

Tracking clouds with a sky camera within a very short horizon below thirty seconds can be a solution to mitigate the effects of sunlight disruptions. A Probability Hypothesis Density (PHD) filter and a Cardinalised Probability Hypothesis Density (CPHD) filter were used on a set of pre-processed sky images. Both filters have been compared with the state-of-the-art methods for performance. It was found that both filters are suitable to perform very-short term irradiance forecasting.

The aim of this thesis is to apply and evaluate potential forecasting models for solar power production, based on data from a photovoltaic facility in Sala, Sweden. The thesis evaluates single step forecasting models as well as multiple step forecasting models, where the three compared models for single step forecasting are persistence, autoregressive integrated moving average (ARIMA) and ARIMAX. ARIMAX is an ARIMA model that also takes exogenous predictors in consideration. In this thesis the evaluated exogenous predictor is wind speed. The two compared multiple step models are multiple step persistence and the Gaussian process (GP). Root mean squared error (RMSE) is used as the measurement of evaluation and thus determining the accuracy of the models. Results show that the ARIMAX models performed most accurate in every simulation of the single step models implementation, which implies that adding the exogenous predictor wind speed increases the accuracy. However, the accuracy only increased by 0.04% at most, which is determined as a minimal amount. Moreover, the results show that the GP model was 3% more accurate than the multiple step persistence; however, the GP model could be further developed by adding more training data or exogenous variables to the model.

Solar and other renewable power sources are becoming an integral part of the electrical grid in the United States. In the Southwest US, solar and wind power plants already serve over 20% of the electrical load during the daytime on sunny days in the Spring. While solar power produces fewer emissions and has a lower carbon footprint than burning fossil fuels, solar power is only generated during the daytime and it is variable due to clouds blocking the sun. Electric utilities that are required to maintain a reliable electricity supply benefit from anticipating the schedule of power output from solar power plants. Forecasting the irradiance reaching the ground, the primary input to a solar power forecast, can help utilities understand and respond to the variability. This dissertation will explore techniques to forecast irradiance that make use of data from a network of sensors deployed throughout Tucson, AZ. The design and deployment of inexpensive sensors used in the network will be described. We will present a forecasting technique that uses data from the sensor network and outperforms a reference persistence forecast for one minute to two hours in the future. We will analyze the errors of this technique in depth and suggest ways to interpret these errors. Then, we will describe a data assimilation technique, optimal interpolation, that combines estimates of irradiance derived from satellite images with data from the sensor network to improve the satellite estimates. These improved satellite estimates form the base of future work that will explore generating forecasts while continuously assimilating new data.

Magíster en Ciencias de la Ingeniería, Mención Eléctrica
En un trabajo previo se propuso un control de mitigación de efecto nube que permitía disminuir los efectos nocivos de la nubosidad parcial sobre parques fotovoltaicos en la frecuencia de sistemas eléctricos de potencia. Esto último sin la necesidad del uso de acumuladores de energía. La estrategia se basa en la operación sub-óptima de los parques (operación en deload) con tal de disponer de reservas de potencia. A pesar que la implementación del sistema nombrado mejoró la frecuencia del sistema de forma significativa en comparación al caso base (sin el sistema de control), la operación en deload de los parques implica una gran cantidad de energía que no se está aprovechando, lo que no se consideró en la metodología. Con tal de mejorar esto, el siguiente trabajo propone un control de mitigación de efecto nube en parques fotovoltaicos de gran escala basado en una herramienta de pronóstico de radiación. Esto último permite disminuir las pérdidas de energía junto con mitigar los efectos de la nubosidad parcial, mediante la determinación de un nivel de deload en los parques fotovoltaicos usando dicho pronóstico. En primer lugar, esta tesis presenta una revisión bibliográfica y discusión del estado del arte de las técnicas de pronóstico en parques fotovoltaicos. Se muestra que la selección de la técnica de pronóstico depende en la información disponible y la ventana de tiempo del pronóstico, es decir, dependerá del caso de estudio. Dicho esto, se propone el uso de una técnica de pronóstico basada en redes neuronales en el Sistema Interconectado del Norte Grande (SING) de Chile. El pronóstico sirve para determinar el nivel de deload en el parque fotovoltaico para los siguientes 10 minutos, en función de una rampa de radiación. Los resultados muestran que la implementación de la técnica de pronóstico no solo mejora la respuesta en frecuencia del sistema, sino que también disminuye las pérdidas energéticas de forma significativa.
Este trabajo fue parcialmente financiado por el Proyecto CONICYT/FONDAP/15110019 "Solar Energy Research Center" SERC-Chile y el Instituto de Sistemas Complejos de Ingeniería (ISCI)

Submitted by Pieter de Jong (pieterj@ufba.br) on 2017-09-06T00:37:40Z No. of bitstreams: 1 DE JONG, Pieter - Forecasting, integration, and storage of renewable energy generation in the Northeast of Brazil.pdf: 4104557 bytes, checksum: 286729716f48be77c37ddb6f4cf2c93d (MD5)
Approved for entry into archive by Vanessa Reis (vanessa.jamile@ufba.br) on 2017-09-06T10:26:26Z (GMT) No. of bitstreams: 1 DE JONG, Pieter - Forecasting, integration, and storage of renewable energy generation in the Northeast of Brazil.pdf: 4104557 bytes, checksum: 286729716f48be77c37ddb6f4cf2c93d (MD5)
Made available in DSpace on 2017-09-06T10:26:26Z (GMT). No. of bitstreams: 1 DE JONG, Pieter - Forecasting, integration, and storage of renewable energy generation in the Northeast of Brazil.pdf: 4104557 bytes, checksum: 286729716f48be77c37ddb6f4cf2c93d (MD5)
CAPES e FAPESB.
As a result of global climate change, during the coming decades less rainfall and higher temperatures are projected for the Brazilian Northeast (NE). Consequently these regional climatic changes could severely impact hydroelectric generation in the NE as well as influence solar and wind power potential. The ongoing drought in the Brazilian NE region has caused hydroelectric generation to decline substantially during the last 5 years and in 2016 hydroelectricity only supplied 25% of the NE’s total demand. In contrast, wind power supplied 30% of demand and is expected to generate 55-60% of the NE’s electricity supply by 2020. Therefore, this paper is focused on both short term forecasting and long-term projections of renewable energy generation and resource availability. It also explores the economic, environmental and technical feasibility of renewable energy integration in the NE region of Brazil. First, the long-term impacts of climate change on the NE region’s hydroelectric and wind energy production are analysed. Particular attention is paid to the long-term projections of annual rainfall and streamflow in the São Francisco basin which could decline by approximately 47% and 80%, respectively, by 2050. On the other hand, wind energy potential is projected to increase substantially during the same period. This thesis also estimates the economic, social, and environmental viability of renewable and non-renewable generation technologies in Brazil. The Levelised Cost of Electricity (LCOE) including externalities is calculated for several different case study power plants, the majority of which are located in the Brazilian NE. It was found that wind power becomes the cheapest generation technology in the NE region, once all externality and transmission line costs are taken into consideration. The LCOE for the entire Northeast’s generation matrix is calculated for various configurations, including scenarios in which hydroelectric generation is restricted due to drought conditions. It was concluded that a generation mix in which wind power replaces all fossil fuel generation by 2020, could feasibly reduce the overall LCOE in the region by approximately 46% and substantially decrease CO2eq emissions. Two different methods are used to examine the limits of integrating high penetrations of variable renewable generation technologies into a power system with a large proportion of hydroelectric capacity. In the first method existing wind generation data from 16 wind farms is extrapolated in time and space, while the second method uses a numerical weather prediction model to simulate future wind energy generation in the NE region. Considering the minimum generation requirements of the São Francisco’s hydroelectric dams, the maximum wind energy penetration in the NE region is estimated to be approximately 50% before significant amounts of energy would need to be curtailed or exported to other Brazilian regions. Finally, this thesis reviews additional literature on energy storage and the impact of large scale variable renewable energy integration on grid stability and power quality. It was found that there are several existing technologies such as power factor and voltage regulation devices that can resolve these issues.
Como consequência da mudança climática global, nas próximas décadas menos precipitação e temperaturas mais altas são projetados para Nordeste (NE) do Brasil. Consequentemente, essas mudanças climáticas regionais podem afetar severamente a geração hidrelétrica no NE, bem como influenciar o potencial de energia solar e eólica. A seca atual nessa região do Brasil fez com que a geração hidrelétrica caísse substancialmente durante os últimos 5 anos e em 2016, as usinas hidrelétricas apenas forneceram 25% da demanda total do NE. Em contraste, a energia eólica forneceu 30% da demanda e deverá gerar 55-60% do fornecimento de energia elétrica do NE até 2020. Portanto, este trabalho está focado tanto na previsão a curto quanto projeções a longo prazo da geração de energia renovável e disponibilidade de recursos. Ele também explora a viabilidade econômica, ambiental e técnica da integração de energias renováveis na região NE. Primeiramente, os impactos de longo prazo das mudanças climáticas na produção hidrelétrica e eólica da região NE são analisados. Especial atenção é dada às projeções de longo prazo de precipitação anual e fluxo na bacia do São Francisco, que podem diminuir em aproximadamente 47% e 80%, respectivamente, até 2050. Por outro lado, prevê-se que o potencial da energia eólica aumente substancialmente durante o mesmo período. Esta tese também estima a viabilidade econômica, social e ambiental das tecnologias de geração renováveis e não-renováveis no Brasil. O custo nivelado de energia elétrica (LCOE), incluindo externalidades, é calculado para diversas usinas de estudo de caso, a maioria localizada no NE. Verificou-se que, a energia eólica se torna a tecnologia de geração mais barata na região NE, uma vez que todos os custos de externalidades e de linhas de transmissão são levados em consideração. O LCOE para a matriz de geração do Nordeste é calculado para várias configurações, incluindo cenários em que a geração hidrelétrica é restrita devido às condições de seca. Concluiu-se que, uma mistura de geração em que a energia eólica substitui toda a geração de combustíveis fósseis até 2020, poderia reduzir o LCOE na região em aproximadamente 46% e diminuir substancialmente as emissões de CO2eq. Dois métodos diferentes são usados para examinar os limites da integração de altas penetrações de tecnologias de geração renovável variáveis em um sistema de energia com uma grande proporção de capacidade hidrelétrica. No primeiro método, dados de geração eólica existentes de 16 parques eólicos são extrapolados no tempo e no espaço, enquanto o segundo método utiliza um modelo de previsão numérica de tempo para simular a futura geração de energia eólica na região NE. Considerando as exigências mínimas de geração das hidrelétricas do São Francisco, estima-se que a penetração máxima de energia eólica na região NE seja de aproximadamente 50% antes que quantidades significativas de energia precisem ser desperdiçadas ou exportadas para outras regiões brasileiras. Finalmente, esta tese examina literatura adicional sobre armazenamento de energia e o impacto da integração de energia renovável variável em larga escala na estabilidade da rede elétrica e na qualidade da energia. Verificou-se que existem várias tecnologias existentes, como dispositivos de regulação de fator de potência e tensão que podem resolver estes problemas.

Due to the rapid growth of grid-tied solar photovoltaic (PV) systems in the generation mix, the distribution grid will face complex operational challenges. High PV penetration can create overvoltages and voltage fluctuations in the network, which are major concerns for the grid operator. Traditional voltage control devices like switched capacitor banks or line voltage regulators can alleviate slow-moving fluctuations, but these devices need to operate more frequently than usual when PV generation fluctuates due to fast cloud movements. Such frequent operations will impact the life expectancy of these voltage control devices. Advanced PV inverter functionalities enable solar PV systems to provide reliable grid support through controlled real injection and/or reactive power compensation. This dissertation proposes a voltage regulation technique to mitigate probable impacts of high PV penetrations on the distribution voltage profile using smart inverter functionalities. A droop-based reactive power compensation method with active power curtailment is proposed, which uses the local voltage regulation at the inverter end. This technique is further augmented with very short-term PV generation forecasts. A hybrid forecasting algorithm is proposed here which is based on measurement-dependent dynamic modeling of PV systems using the Kalman Filter theory. Physical modeling of the PV system is utilized by this forecasting algorithm. Because of the rise in distributed PV systems, modeling of geographic dispersion is also addressed under PV system modeling. The proposed voltage regulation method is coordinated with existing voltage regulator operations to reduce required number of tap-change operations. Control settings of the voltage regulators are adjusted to achieve minimal number of tap-change operations within a predefined time window. Finally, integration of energy storage is studied to highlight the value of the proposed voltage regulation technique vis-à-vis increased solar energy use.
Ph. D.

This thesis presents different optimization and forecasting models, with the focus on energy markets and renewable energy sources. The analysis approach is related to models for wind and solar power forecasts and those for electricity prices forecasts. The first study explores a Principal Component Analysis in combination with two post-processing techniques for the prediction of wind power and of solar irradiance produced over two large areas. The Principal Component Analysis is applied to reduce the datasets dimension. A Neural Network and an Analog Ensemble post-processing are then applied on the PCA output to obtain the final forecasts. The study shows that combining PCA with these post-processing techniques leads to better results when compared to the implementation without the PCA reduction. The second work explores two different techniques for the prediction of the Italian day-ahead electricity market prices. The predicted Italian prices are the zonal prices and the uniform purchase price (Prezzo Unico Nazionale or PUN). The study is conducted using hourly data of the prices to be predicted and a large set of variables used as predictors (i.e. historical prices, forecast load, wind and solar power forecasts, expected plenty or shortage of hydroelectric production, net transfer capacity available at the interconnections and the gas prices). A Neural Network and a Support Vector Regression are applied on the different predictors to obtain the final forecasts. Different predictors’ combinations are analysed to find the best forecast. The results show that the best configuration is obtained using all the predictors together and applying the Neural Network to find the forecasted prices.

With the potential to sustain the world’s energy needs, solar energy plays a major role for the renewable energy transition. However, inherent problems exist in solar energy forecasting, a very important tool for power plant operators that allows an efficient energy management and dispatch operations in the electric grid. In particular, concentrating solar power (CSP) systems, which rely on direct normal irradiance (DNI) and its high variability, which links uncertainty to the electrical energy outputs of CSP plants. The main atmospheric factors that influence DNI variation at surface are clouds and aerosols, which are misrepresented by current numerical weather prediction models. To provide accurate predictions of DNI for efficient CSP operations, particularly during periods of direct solar intermittency, the solar resource needs to be well characterized. Solution to this problem is still one of today’s challenges in solar forecasting. This thesis makes use of short-term forecasts of DNI from the Integrated Forecasting System (IFS), the global model of the European Centre for Medium-Range Weather Forecasts (ECMWF), together with ground-based measurements in Alentejo region (southern Portugal). The evaluation of the solar resource in the region is based on the IFS predictions, as well as the prediction of energy production outputs from different CSP systems through the System Advisor Model (SAM) power plant simulator, in which the results are compared with local measured data. To improve the role that DNI forecasting has in CSP power plants, several post-processing techniques are used for the correction of hour and day-ahead values of DNI. Different operational strategies are discussed and proposed according to the obtained results; Resumo: Previsão de Curto Tempo de Radiação Normal Directa Através de Modelos Numéricos de Previsão do Tempo no Alentejo (Sul de Portugal): Implicações para as Tecnologias de Concentração Solar Com potencial para assegurar as necessidades energéticas do mundo, a energia solar desempenha um papel importante na transição energética renovável. Contudo, existem problemas inerentes na previsão de energia solar, uma ferramenta muito importante para os operadores de centrais eléctricas que permite uma gestão energética mais eficiente e operações de distribuição da mesma na rede eléctrica. Em particular, os sistemas de concentração de energia solar (CSP), que dependem da radiação normal directa (DNI) e da sua elevada variabilidade, atribuindo incerteza à geração de energia eléctrica resultantes de centrais CSP. Para fornecer previsões precisas para operações CSP eficientes, particularmente durante períodos de intermitência solar directa, o DNI precisa de ser bem caracterizado. Os principais factores atmosféricos que influenciam a variação de DNI à superfície são as nuvens e os aerossóis, que não são representados realisticamente pelos actuais modelos numéricos de previsão do tempo. A solução para este problema é ainda hoje em dia um desafio em previsão solar. Esta tese faz uso das previsões de curto-período de DNI do Integrated Forecasting System (IFS), modelo global do European Centre for Medium-Range Weather Forecasts (ECMWF), em conjunto com medidas à superfície na região do Alentejo (sul de Portugal). A avaliação do recurso solar na região é efectuada com base em previsões do IFS, tal como a previsão de outputs de produção energética de diferentes sistemas CSP através do simulador de centrais eléctricas System Advisor Model (SAM), onde os resultados são comparados com os obtidos com medidas meteorológicas locais. Para melhorar o papel que a previsão de DNI tem em centrais CSP, várias técnicas de pós-processamento são efectuadas para a correção de valores de DNI para a hora e dia seguinte. Diferentes estratégias de operação são discutidas e propostas de acordo com resultados obtidos.

碩士
正修科技大學
電機工程研究所
105
Following the hydroelectric power generation and thermal power generation, nuclear has nowadays become one of the major options for the electricity generation since been invented. Though the nuclear has been considered as the most effective solution for the electricity generation, it also bears the odium for the high risk of the mass destruction if any accident occurred. The tendency toward "Nuclear Free" or "Anti-Nuclear" has triggered a heated debate after the Fukushima nuclear accident in Japan. Taiwan, as an island country, has the extremely high needs for importing the various fuel of the electricity generation in the absence of water resource and natural mineral. Under such circumstance, the development of solar power has answered the call upon the power generation in every country in the world. In the period of 2000 to 2015, the sum of yearly solar power generated in Taiwan has the growth from 1,200 kWh to 850,300 kWh. In addition, the installed capacity of solar power in Taiwan has the growth of 606 MW during 2009 to 2014 after the "Renewable Energy Development Act" be taken in to effect by "Bureau of Energy, Ministry of Economic Affairs, R.O.C." With the result mentioned above, it is showed that the solar power is becoming as the major option in electricity generation in Taiwan. Though solar power has been considered as alternative power resource in the parallel system, it will also create the impact for the power system due to the unstable power feedback result from the various environment factors. Except the irradiance, the atmospheric environment factors, such as temperature and humidity, shall be also considered as the root cause in the solar power generation. This research is aimed to analyse the relevance among solar power generation, temptation and humidity, and further provide an effective procedure for forecasting the solar power generation in short term. With the auxiliary feedback received in this solution, we expect the power system could maintain the stability by predicting the change of the immediate loading and adjustment of reserve margin.

The production of solar energy is undoubtedly dependent on weather conditions and the time of day in which it occurs. Therefore, this source of energy being extremely variable and difficult to predict with high accuracy after many studies, the need for a new approach surges in order to increase it. With this project, it is proposed the use of images as input for deep learning structures with the purpose to achieve automatic feature extraction that can be replicated to other data sets. The images are generated from a grid of numerical weather predictions of, particularly, surface downelling shortwave flux and cloud cover at different levels. In this thesis, several neural network models with be analysed according to its input data: numerical, numerical conjugated with images and images. The models with numerical input data will serve as reference for comparison and evaluation of the added value brought by the images to the forecasts, since the reference's predictions used variables obtained by manual feature extraction.

The production of solar energy is undoubtedly dependent on weather conditions and the time of day in which it occurs. Therefore, this source of energy being extremely variable and difficult to predict with high accuracy after many studies, the need for a new approach surges in order to increase it. With this project, it is proposed the use of images as input for deep learning structures with the purpose to achieve automatic feature extraction that can be replicated to other data sets. The images are generated from a grid of numerical weather predictions of, particularly, surface downelling shortwave flux and cloud cover at different levels. In this thesis, several neural network models with be analysed according to its input data: numerical, numerical conjugated with images and images. The models with numerical input data will serve as reference for comparison and evaluation of the added value brought by the images to the forecasts, since the reference's predictions used variables obtained by manual feature extraction.

碩士
大同大學
電機工程學系(所)
107
Solar photovoltaic (PV) generation has received great attention in recent years due to the promotion of green environment awareness. Accurate forecasting of solar power benefits the preparation for switching other renewable energies into the power grids when PV power becomes low. In particular, the harmful consequence from the large peak and off-peak gaps of the so-called “duck curve” for PV power can be mitigated. Motivated by the fact that the contingency reserve typically requires thirty to sixty minutes to start up, we mainly focus on the PV power prediction at the hourly level. Among numerous studies in the literature dealing with solar power forecasting via various machine learning methods, the application of long short-term memory (LSTM) on hourly PV power prediction has been recently proposed to capture both hourly patterns in a day and seasonal patterns across days. For hourly prediction, however, we argue that the contribution of seasonal factors might be marginal since the correlation of meteorological information between days is much higher than between years. In this paper, we consider the LSTM-based hourly PV power prediction with the daily factor (24 hours) instead of the seasonal factor (day and month) and improve the pattern learning by selecting the highly-effective features toward the different number of time steps. This design is to enhance the information of time-dependency between each set of training input which is crucial for hourly prediction. By using the real-world data, the experimental results show that the accuracy of hourly PV power prediction can improve from 92–95% to 94–95.5% in terms of coefficient of determination (R2) whilst using the same dataset.

The Sun is an ultimate source of energy for all the objects living on Earth and solar radiation estimation is utmost important as this radiation serves as a primary energy source of conversion for the photovoltaic panels and the solar thermal power plants. This solar radiation is not constant in every region but it depends on various climatological parameters like temperature, wind-speed and many more, so there is intermittency in its behavior which results in changes in the electrical energy production. The above few statements reveal the necessity of predicting solar radiation. Physical method, Statistical method, Hybrid method are namely the methods put forward by the researchers around the globe for the purpose of forecasting Global Solar Radiation. The time-series forecasting method (ARIMA, ARX, AR models) are the oldest of all the methods thus there are countless research materials on them but in the recent years forecasting with the help of Machine Learning techniques (k-NN, SVM, Random Forest) have gained popularity over the time series method of forecasting. Today, the world is moving in a fast pace with regards to energy and power. It is something which is taken for granted. In order to suffice this demand of energy globally we are mostly dependent on non-renewable sources of energy which with its high-time usage is soon going to deplete and is a major cause of global warming. So, the world is looking for an environmentally friendly energy resources, the solar energy proves to be an important clean energy source. Entire requirements of human population can easily be met by the amount of solar energy that falls on the earth’s surface every hour. Thus, it becomes necessary for the industries to switch from traditional sources to solar energy as its primary source of energy and this requires an accurate prediction of solar power. The solar power forecasting not only determines the size of the operating reserves for generation-load balance but also reduces the operating cost thereby improving the reliability of the grid. In this research work, the 3 Machine Learning algorithms i.e., K-Nearest Neighbors, Support Vector Machine, Random Forest are used to build models for accurately predicting solar power of Kurnool region of Karnataka (India).At the tail end of the paper the evaluation metrics RMSE, MAPE, MAE etc. are also calculated to measure the accuracy of each algorithm and a significant comparison is made among them to come up on a conclusion that Random Forest method gave the best overall performance with RMSE=0.6157, MAPE=2.6421, MAE=0.4903 and R2=0.6517

碩士
國立臺灣師範大學
工業教育學系
104
Because solar irradiance is susceptible to clouds and substances in the air, the solar photovoltaic cannot produce stable power output. The power output of a photovoltaic module is influenced immediately when the module is sheltered from the clouds. Besides, the material of solar cell, air temperature, module’s position and orientation also affect the power output of the photovoltaic module. The main goal of the thesis is to develop the solar power forecasting with 24 hours ahead by applying back-propagation neural network technology. Some different combination inputs of the back-propagation neural network are proposed and their forecasting performances are evaluated. Moreover, comparison results in Taichung solar farm are given. As a result, the better performance is achieved by the inputs with combination of future factors.

碩士
聖約翰科技大學
電機工程系碩士班
103
This thesis proposes an adaptive network-based fuzzy inference system (ANFIS) based forecasting method for short-term wind power and solar power forecasting. An accurate forecasting method for power generation of the wind energy conversion system (WECS) and the photovoltaic (PV) system is urgent needed under the relevant issues associated with the high penetration of wind and solar power in the electricity system. To demonstrate the effectiveness of the proposed forecasting method， the method is tested on the practical information of wind power generation of a 2.3 MW WECS installed on the Taichung coast of Taiwan， and the practical information of solar power generation of a 9 kW PV system installed on the St. John’s University of Taiwan. Comparisons of forecasting performance are made with the persistence method (PM)， back-propagation neural network (BPN) forecasting methods. Good agreements between the realistic values and forecasting values are obtained; the test results show the proposed forecasting method achieves better forecasting performance.

碩士
國立臺灣大學
資訊工程學研究所
105
In recent years, solar energy has become a significant field of research across the globe because of the growing demand for renewable energy and its promising potential in sustainability aspects. Therefore, with the increasing integration of photovoltaic (PV) systems to the electrical grid, reliable prediction of the expected production output of PV systems is gaining importance as a basis for management and operation strategies. However, power production of PV systems is highly variable due to its dependence on solar radiance, meteorological conditions and other external factors. Currently, most studies are unable to predict the solar power production at multiple future time points and never analyze the influence of the possible factors on the solar power production but simply feed all features without considering their properties. Therefore, this paper provides a holistic comparison among all factors (e.g., solar radiance, meteorology) affecting the solar power production and also, a multimodal and end-to-end neural networks model is proposed to simultaneously predict the solar power production at multiple future time points by multitask learning. Finally, with multitask learning on heterogeneous (and multimodal) data, the proposed method achieves the lowest error rates (11.83\% for 5-minute prediction) compared to the state of the art.

碩士
國立中興大學
資訊管理學系所
105
Renewable energy such as solar power is becoming increasingly important as the consumption of non-renewable energy, but how to decide on the location of the solar energy plant? Though we can make the decision from the measurement of some high-precision instruments, it costs a lot. Apart from that, we thus can use prediction model for judging whether a location is suitable for the development of solar energy. In this study, we use the solar energy data and the data of the Central Weather Bureau, and the data interval is from January 1, 2015 to December 29, 2015, then through the merger of data, screening variables, and the establishment of multiple prediction models. Then we use neural networks, support vector machines and other algorithms. Repeat training model, and adjust the hidden layer of nerve Meta-number and kernel function, to find out the best model for predicting the amount of solar power in Taiwan. The results show that the use of neural networks to predict the amount of solar power in Taiwan is more accurate. Besides, from original sixteen variables, we remove the variables which has low impact for prediction or is obtained difficult variables, the last ten variables can contribute to the best prediction in model 7 whose MAPE is 12.25%. Furthermore, we are looking forward to Taiwan to have more solar power data, other external factors and meteorological data, and thus we have a more accurate model and predict the results.

碩士
國立臺灣大學
電機工程學研究所
107
The development plan of the PV power plant in Taiwan with high penetration makes the grid operator has to prepare the strategies to mitigate its intermittency behaviour. The power grid must be more flexible to receive the fluctuated power from PV. One of the economical ways is to conduct a solar power forecasting. The solar power forecasting is made using artificial neural networks in this thesis. The networks are trained using backpropagation and extreme learning machine. The extreme learning machine was used due to its advantages in accuracy and computational time over backpropagation neural networks. The persistence model is used as the reference for the performance index. It can also be the input of the combined forecasting model to improve accuracy. Later on, the month-based segmentation forecasting was investigated to accommodate the seasonal variation of generated PV power output. At the end of the research, it is found out that the segmented solar power forecasting has a better performance than the forecasting with only one model for the whole year.

The vanishing conventional energy sources and global warming drive the world for the power generation from renewable energy sources. The main renewable sources namely solar power and wind power are uncertain and intermittent in nature. Wind & solar photo voltaic (PV) power forecasting with good accuracy promise the power sector for large scale integrations of wind & solar PV power generations into the grid. In the context of smart grid and deregulated electricity market, price forecasting is a challenging job for researchers. A rigorous literature review of wind power forecasting, solar PV power forecasting and price forecasting is conducted with focus on various statistical & learning forecasting methods. The data is collected from Belgium wind farms, US wind farms and Indian wind farms and Indian photo voltaic plants. The dependency of wind power generation and solar power generation is analyzed with the computation of correlation factors. Nonlinear autoregressive with external input (NARX) model is implemented to forecast wind power generation of Belgium wind farms by using historical data of wind speed and wind power. Further, NARX model is also used to forecast wind speed for US wind farms from the input data of wind direction, temperature and air density. Wind speed is predicted with good accuracy and minimum MAPE is 2.3%. The research work is continued to improve short term wind power forecasting accuracy by designing generalized regression neural network (GRNN) and radial basis function neural network (RBFN). A hybrid network of GRNN &RBFN is designed with parallel topology to forecast wind power for improved accuracy. Reliability of forecasting models is analyzed with the computation of confidence intervals on MAPE. As support vector machine (SVM) is very good at classification and regression analysis, in this work the support vector regression (SVR) model with tuned parameters is used to xiii forecast wind power generation and solar PV power generation. To achieve better accuracy and to retain the benefits of individual models, a hybrid approach K-means clustering based artificial neural network- particle swarm optimization (ANN-PSO) model is designed and proposed for solar PV power forecasting. In the context of smart grid, the uncertainty in wind & solar PV power generations increases the volatility of electricity price. A hybrid approach of K-means clustering based long short term memory (LSTM) network is proposed for short term electricity price forecasting of Austria by considering wind power generation in the market. The proposed model shows highest accuracy in prediction when compared against feed forward neural network-particle swarm optimization (FNN-PSO) and SVR models. In hour ahead price forecasting with the consideration of wind & solar PV power generations, bootstrap aggregation of ensemble model (proposed model) has outperformed with significant reduction in error. As renewable energy integration to the power grid is enhancing day by day, it becomes pertinent to introduce new market models to operate the renewable energy (RE) enabled restructured electricity market. For such an RE enabled Indian electricity market, seven various market models are developed and proposed along with their salient features. An operating mechanism for future RE enabled Indian electricity market is also proposed based upon the developed models.

碩士
國立高雄應用科技大學
電機工程系博碩士班
106
Based on the existed solar power generation data in Taiwan, this thesis applies parallel Elman Neural Network associated with solar radiation and system conversion efficiency as parameters to construct a Taiwan solar energy forecasting model. The forecasting accuracy is verified by the information of photovoltaic power station with different regions and sizes. As well as the prediction model is estimated by K-means and inverse distance weighting skills to improve the solar power generation in various regions of Taiwan. The reliability of the estimation results is confirmed by the photovoltaic power station in Miaoli and Pingtung areas. The estimation results of power generation in these areas help the Taipower dispatching center to accurately grasp the trend of solar power generation in various regions, and at the same time coordinate with the fossil power and hydraulic power to meet accurately load demand. The proposed model will support the benefit to power dispatch for larger scale intermittent unstable solar power generation in the future.

MSc (Statistics)
Department of Statistics
This study discusses an application of partially linear additive quantile regression models in predicting medium-term global solar irradiance using data from Tellerie radiometric station in South Africa for the period August 2009 to April 2010. Variables are selected using a least absolute shrinkage and selection operator (Lasso) via hierarchical interactions and the parameters of the developed models are estimated using the Barrodale and Roberts's algorithm. The best models are selected based on the Akaike information criterion (AIC), Bayesian information criterion (BIC), adjusted R squared (AdjR2) and generalised cross validation (GCV). The accuracy of the forecasts is evaluated using mean absolute error (MAE) and root mean square errors (RMSE). To improve the accuracy of forecasts, a convex forecast combination algorithm where the average loss su ered by the models is based on the pinball loss function is used. A second forecast combination method which is quantile regression averaging (QRA) is also used. The best set of forecasts is selected based on the prediction interval coverage probability (PICP), prediction interval normalised average width (PINAW) and prediction interval normalised average deviation (PINAD). The results show that QRA is the best model since it produces robust prediction intervals than other models. The percentage improvement is calculated and the results demonstrate that QRA model over GAM with interactions yields a small improvement whereas QRA over a convex forecast combination model yields a higher percentage improvement. A major contribution of this dissertation is the inclusion of a non-linear trend variable and the extension of forecast combination models to include the QRA.
NRF

碩士
國立雲林科技大學
資訊管理系
104
The rising importance of solar energy comes along with the attention of solar power forecasting. Accurate forecasting of solar power output is crucial to power distribution and plant management. This study aims to compare performances of different forecasting algorithms coming from machine learning. Comparisons also make on using multiple types of datasets, different sets of input features, different parameter settings, and different solar power plants. In particular, we investigate the influences between the meteorological factors and power outputs, sun path variables and power outputs. At the end, a prototype of solar power forecasting system is constructed and will be deployed on several field plants. In this research, two forecasting algorithms, multiple layer perceptron neural network (MLPNN) and k-nearest neighbors (KNN) are applied to develop the forecasting models. The experimental results indicate that radiation is the most crucial factor for the power output forecasting and MLPNN outperforms KNN algorithm. The performance of KNN can be improved if the predicated values of radiation are stable and fewer outliers.

This thesis deals with quantifying the merits of Distributed Maximum Power Point Tracking (DMPPT), as well as providing solutions to achieve DMPPT in PV systems. A general method based on 3D modeling is developed to determine the energy yield of PV installations exploiting different levels of DMPPT granularity. Sub-string-level DMPPT is shown to have up to 30% more annual energy yield than panel-level DMPPT. A Multi-Input-Single-Output (MISO) dc-dc converter is proposed to achieve DMPPT in parallel-connected applications. A digital current-mode controller is used to operate the MISO converter in pseudo-CCM mode. For series-connected applications, the virtualparallel concept is introduced to utilize the robustness of the parallel connection. This concept is demonstrated on a three-phase boost converter. The topology offers reduced output voltage ripple under shading which increases the life-time of the output capacitor. The prototypes yield output power benefits of up to 46% and 20% for the tested shading conditions.

In the last few years, several countries have accomplished their determined renewable energy targets to achieve their future energy requirements with the foremost aim to encourage sustainable growth with reduced emissions, mainly through the implementation of wind and solar energy. Wind and solar energy is critically important for the social and economic growth of any country. Moreover, reliable and precise wind and solar power prediction is crucial for the dispatch, unit commitment, and stable functioning of power systems. This makes it easier for grid operators of the power system to support uniform power distribution, reduce energy loses, and optimize power output. Consequently, the integration of wind and solar power globally relies on correct wind and solar power forecasting. Current studies typically adopt machine learning algorithms (ML). The foremost contribution of this research is short-term wind power forecasting on the basis of the historical values of wind speed, wind direction, and wind power by using ML algorithms. In this study, regression algorithms such as random forest, k-nearest neighbor (k-NN), gradient boosting machine (GBM), decision tree, and extra tree regression are employed to enhance the forecasting accuracy for wind power production for a Turkish wind farm situated in the west of Turkey. Polar curves have been plotted and the impacts of input variables such as the wind speed and direction on wind energy generation is examined. Scatter curves depicting the relationships between the wind speed and the produced turbine power are plotted for all of the methods here and the predicted average wind power is compared with the real average power from a turbine with the help of the plotted error curves. The second contribution of this research is short-term solar power forecasting on the basis of the historical values of ambient temperature, irradiation, module temperature and solar power by using ML algorithms. In this study, regression algorithms such as random forest (RF) and k-nearest neighbor (k-NN) regression algorithms are employed to enhance the forecasting vii accuracy for solar power production for a Qassim University, KSA. The performance of all algorithms were estimated based on the various statistical indicators. As renewable energy sources (RES) provide intermittent power and are not available 24 hours a day, it is vital to build hybrid models based on RES to provide an uninterrupted, sustainable, eco-friendly, and cost-efficient power supply. The current research is also devoted to the development and design of an optimal hybrid model using locally accessible RES for selected locations. The evaluation of the potential of locally available RES for selected sites in Uttar Pradesh, India, is carried out to develop the hybrid model. To fulfil the energy demand of the selected site, a hybrid model was constructed using the Hybrid Optimization Model for Electrical Renewable (HOMER) software based on the feasibility analysis of RES at the selected site. To create a hybrid model, the electrical load demand for the specified location is evaluated while taking seasonal fluctuations, current and future power requirements, everyday hourly consumption patterns, living standards, and so on into account. The primary goal of this study is to develop an economic and optimal hybrid PV/Biogas configuration for power production for rural common facilities including one Primary school, Junior school and Panchayat Ghar buildings of Sarai Jairam village in Uttar Pradesh, India. The PV/biogas hybrid configuration was designed utilizing the Hybrid Optimization Model for Electric Renewable (HOMER) and techno-economic analysis is carried out to fulfill the load requirements. The HOMER analysis produced a solution that included total net present cost (NPC) and cost of electricity (COE), and these results were then further improved using sensitivity analysis. Based on the NPC and COE, this analysis evaluates the system performance and demonstrates that it is techno-economically feasible. In addition, for maximizing the solar power generated from solar photovoltaic system (SPV), the optimization of space and orientation of solar PV system are also done.