Improved modelling of microgrid distributed energy resources with machine learning algorithms
- Autores: Miguel Carpintero Rentería
- Directores de la Tesis: David Santos Martin (dir. tes.)
- Lectura: En la Universidad Carlos III de Madrid ( España ) en 2021
- Idioma: español
- Tribunal Calificador de la Tesis: Carlos Veganzones Nicolás (presid.), Mª Ángeles Moreno López de Saá (secret.), Athanasios Kolios (voc.)
- Programa de doctorado: Programa de Doctorado en Ingeniería Eléctrica, Electrónica y Automática por la Universidad Carlos III de Madrid
- Materias:
- Enlaces
- Tesis en acceso abierto en: e-Archivo
- Resumen
Distributed energy resources (DERs) are electrical power generation technologies or storage systems connected to low- or medium-voltage networks. These types of technologies, generally composed of renewable systems, have experienced a significant growth due to increased energy demand and the reduced costs of some renewable technologies. Nevertheless, increasing the penetration of different generation technologies adds an extra layer of complexity to the operation, control and assessment of the grid behaviour. Structures such as microgrids (MGs) may provide an ecosystem to operate distributed generation (DG) with resiliency, efficiency and flexibility when connected to a bigger grid or while working in isolation, like island grid systems. With high renewable penetration levels and storage systems connected to the grid, an adequate modelling of each power generation unit is needed. Even though it is important to study the operation and control of the interconnected systems, a solid performance of each individual model is crucial. These individual models can be used in dynamic studies, unit commitment or planning software, among other uses.
Given the advancements in electronics, many data acquisition devices have been embedded with DERs and the grid. These sensors generate massive amounts of information that must be handled and generally administered by the supervisory control and data acquisition system. Data scientists may use this information to better understand DERs behaviour so that more optimised grid systems are built. Within the scope of data science, ML algorithms have become increasingly popular in many different domains. Electrical engineering is no exception given the difficulty of depicting and modelling complex data or behaviours. Hence, the use of ML techniques to model individual DERs becomes a natural and meaningful path to explore.
This thesis starts by studying the MGs through an extensive literature review. Then, after exploring a wide range of DERs associated to real world case MGs, an ML search model is proposed to perform infrastructure examples of MGs based on segment of operation and location. Also, a statistical analysis is developed. Then, the principal renewable DERs associated to the previous study are selected to perform different ML models and validate the results with real data and classical modelling approaches when possible. A brief description of each thesis’ study is detailed below.
Regarding the MG review, within a DG system, MGs are an alternative approach that may provide both resiliency and efficiency benefits. In this review, an analysis of both research and industrial documents was done. In order to establish a solid foundation of the MGs concept, a comparison of various definitions written by distinguished authors has been made. Segmenting the information of MGs into layers facilitates its analysis, search, and comparison. Therefore, this chapter continuous with a layer approach from other studies and incorporates the concept of the environment as a key element that has a high impact on the MG functional structure. The environment differentiates two types of layers, external and internal.
Within the external layers, the policies and standards layer interact and affects not only the other layers but the deployment of MGs themselves. Despite that the use of DG standards has been adopted, new particular MG standards and policies are emerging. Regarding the business layer, it has been studied from two points of view, one related to different business models proposed by some authors and the other one showing how policies are very close related with this layer. The last external layer is how the environment and the climate differences have a key role into the selection of components and the approach followed in the other layers such as the operation and control strategies.
Regarding the internal layers analysed, the infrastructure perspective is done again from two points of view, one regarding the electrical components involved and the other one regarding the architecture and strategy used to combine all these different components. The last layer analysed is also considered as a distinctive and key factor by many authors. Widely studied, the operations and control layer is disaggregated into two—the strict control strategies analysed by the authors and the management and operation techniques used.
Throughout the analysed literature there are multiple definitions of an MG. Just two of those definitions are presented in a technical standard, but all of them share three common factors: (a) islanding capabilities, (b) clearly defined boundaries, and (c) control to operate both resources and loads.
After conducting the literature review, a compendium about MGs was performed. In this project, detailed information about the infrastructure layer in MG projects is available, therefore the aggregated information based on 1,618 MGs was summarised into different tables and analysed based on various parameters. Two MG infrastructure model creation tools were developed, by contrast to previous studies that dealt with fewer number of MGs and no searching tool. By aggregating and representing all the collected data, the aim was to establish a guiding principle for researchers to perceive how real MGs are being deployed around the world.
With the aggregated information of the compendium, an ML tool shared as a MATLAB script with a classification and regression learner and a flowchart guideline are proposed. The ML learning model is similar to a searching tool and generates MG infrastructure models using the latitude and the segment as inputs providing the generation technologies and power capacity based on the compendium data. Those generation technologies can be also be provided by the user and the algorithm provides the power associated to each technology.
While both models can be adapted to the researcher’s needs; there is a difference in the outcome. The model creation guideline produces more generic MG infrastructure models than the ML algorithms due to the resolution in the latitude presented in the tables and the correlations made among the data by the algorithm. Actually, the ML tool latitude precision is close to 2° instead of the 15 − 30° from the guideline.
The principal renewable DERs associated to MGs are photovoltaic (PV) systems, batteries and wind energy conversion systems (WECS) respectively. First, multiple PV models divided in two studies are presented in the thesis, then an ML model to estimate the state of charge (SOC) of a battery is shown. Lastly, different models associated to WECS are proposed in two different projects.
The first project associated to PV systems is performed due to the increase penetration of this technology in the electrical network. Despite there is a large variety of PV models for power production estimation. The use of ML algorithms in this application is yet to be explored. In this study, a comparison between statistical and deterministic approaches is presented. The statistical models proposed use ML methodologies with neural network algorithms and automatic strategies to clean the data. To validate the model a comparison with two different deterministic models in different situations and locations is analysed.
After developing the study, it can be concluded that neural network (NN) models accurately estimate PV power when trained and tested with data from the same location. This fact makes ML methodologies a very good performance ratio tool for a PV power station when using their own SCADA data. In contrast, the NN model shows worse results when training the algorithm with one location and testing it in a different one. Nevertheless, this provides an opportunity to make future studies with data from more locations, climates or even different types of ML algorithms. After the study, a reassessment of the generalisation capabilities to estimate the PV power output in different locations with the ML model should be done. However, this is a difficult task due to the great amount of data with an adequate resolution required. This study also shows how ML techniques have been successfully applied to obtain accurate results by only using SCADA data and features extracted out of it. This makes the ML methodologies a promising tool when the technical specification of components such as the inverter, type of PV panels among some others are missing.
The second study of PV systems is referred to modelling the annual optimum tilt angle for fixed solar PV arrays or solar collectors, in any location of the world. The optimum tilt angle that maximises the annual energy yield can be easily calculated in the absence of meteorological data and simulation software tools. The proposed models are calculated using global horizontal radiation data collected from 2,603 sites across the world. In the process, well-established submodels have been selected to estimate the hourly irradiance on any possible inclined surface, and its corresponding annual energy yield. After selecting the optimum angle for each location, through a regression analysis, a mathematical model that calculates annual optimum angles as a function of latitude has been developed. Furthermore, regression techniques such as NN and decision trees (DT) have been compared with the polynomial models. The results are analysed, validated, and compared with previous research proposals proving the good performance of the proposed models.
It has been found that due to the lack of access to meteorological data in many locations around the world, or to the small size of the installation, it is convenient for designers or users of solar collectors or PV systems to have access to a mathematical model for determining their optimal orientation. Quadratic and cubic forms of the model proposed are provided, with the latter showing slightly better performance.
The polynomial models have been extensively validated against other existing regression models, as well as published optimum tilt angle calculations and measurements. In all cases the models show consistent performance, and good agreement with established results. It was shown, however, that significant differences may be observed when comparing the calculated optimum tilt angle to other models fitted using data covering a short period of time, or a small number of sites. The results of the proposed models are most similar to those which have also been constructed using large datasets. This highlights the importance of using representative irradiance time series from a large number of different sites in the construction of such general models.
While more complex regression methods were also investigated for calculating the optimum tilt angle, including NN and DTs, they only produced a minor improvement in performance. Hence, for the sake of easy implementation, the authors suggest the use of either the quadratic or cubic polynomial models. These models are believed to be the most accurate models of their type to date, due to the size and quality of the datasets used.
After analysing the PV systems, the second most used DER in MGs are the batteries. The available capacity of a battery, called the SOC, is a fundamental characteristic for energy storage applications or electric vehicles. In order to model the SOC of a lithium-ion battery using data-driven techniques, complex algorithms should be used so the dynamic behaviours of the battery are captured. In this study, an ensemble DT called gradient boosting tree was fitted with the information extracted from different experiments based on dynamic and constant discharge profiles at different temperatures and implemented in a laboratory. While analysing the data from the experiments, some tests exhibited unstable behaviour, usually when the battery was working at threshold temperatures and currents. This stochastic operation was likely produced because of the state of health (with a capacity below the theoretical life expectancy) and the chemical degradation of the battery. Regarding the ML model, ensemble gradient boosting trees have not been widely studied in the literature, but according to these experiments, this type of model provided solid results. Due to the state of health of the battery, algorithms and strategies from the literature were tested with inconsistent results and none of them provided low estimation errors with low variability as did the gradient boosting model. From the 28 experiments, only three had generalised noise, but the tendency of the predicted function always followed the true SOC. To avoid dealing with stochastic and unstable behaviour, it is highly recommended that future research performs experiments with a new battery while controlling the degradation with the time. Applying the degradation feature as an input of the ML algorithm for better estimation results is also recommended.
The first WECS study was regarding the power coefficient parameter, which represents the aerodynamic wind turbine efficiency. Since the 1980s, several equations have been used in the literature to study the power coefficient as a function of the tip speed ratio and the pitch angle. In this study, these equations are reviewed and compared. A corrected blade element momentum algorithm is used to generate three sets of data representing different ranges of wind turbine, going from 2 to 10 MW. With this information, two power coefficient models are proposed and shared. One model is based on a polynomial fitting, whereas the other is based on NN techniques. Both were trained with the blade element momentum model output data and showed good behaviour for all operating ranges.
The errors caused on many occasions by this type of approximation can have a great impact in dynamic and transient studies. In this study, both models proposed obtained between 55% and 60% lower error rates than the best numerical approximations found in the literature within this test set. Nevertheless, it would be interesting in the future to study more airfoils over a wider variety of wind turbines to further enhance the NN’s and polynomial fitting performance. The aim of developing the models and sharing them was to obtain good performance when calculating Cp in different types of wind turbines so the models become as universal as possible. These kind of models can be used for static analysis such as power production. Reducing the error rate in the power coefficient parameter may have a great impact on many WECS studies, such as those treating static, dynamic and transient behaviours.
The polynomial model has the advantage of having an easy implementation, with good performance and reduced computational cost. The NN model has better performance, with the possibility to retrain it or use it with a transfer learning approach so researchers can make more robust power coefficient algorithms with less input data.
The last WECS study is referred to 3 different ML models and an automatic filtering algorithm. Copulas models provide a flexible approach to modelling joint probability distributions. Hence, alongside default cleaning techniques copula models can be used to remove outliers. When fitting the copula model to the dataset, different types of outliers can be automatically filtered with this technique. The three case study models proposed show that simple NN architectures are able to accurately model the power delivered by WECS improved by an automatically filtered dataset.
This thesis has contributed to the literature by improving the models of the principal DERs associated with MG systems. These technologies are PV, energy storage and WECS. The ML models proposed can be applied either as individual models for performance assessment of each distributed energy resource or as complementary tools for dynamic or static studies, unit commitment or planning software, among some others. It is time to fully compare traditional models made of physical, electrical or chemical algorithms with data-driven approaches based on ML. Data-driven techniques may provide more custom solutions to improve the performance of models by simulating the behaviour of concrete systems and locations. In this thesis, as previously stated, only a few DERs have been modelled and validated. Thus, given these positive results, future work on this subject should be done.
