Improved modelling of microgrid distributed energy resources with machine learning algorithms

Abstract

Renewable energy technologies are being increasingly adopted in many countries around the world. However, the deployment of these power generation systems is becoming more diverse than ever, from small generation units in individual houses to massive power production plants. In that spectrum, distributed energy resources (DERs) cover systems from the low- to the middle-power ranges. The operation, control and assessment of these technologies is becoming more complex, and structures such as microgrids (MGs) may provide a suitable ecosystem to manage them. The beginning of this thesis covers the fundamentals of MG systems. A review was conducted by analysing the MG in a layer perspective, where each layer corresponded to a topic such as operation, business or standards, among others. The advancements in electronics, computer power and storage capability have created a paradigm in which massive amounts of data are generated and computed. The electrical sector has introduced many data acquisition technologies to assess the grid and its components. Classical modelling approaches have applied physical, chemical or electrical algorithms to model the behaviours of DERs. Nevertheless, with the extensive amount of information at our disposal, data-driven techniques such as machine learning (ML) may provide more individualised models to simulate the behaviour of these power generation technologies with the particularities of both their components and their location. Following the MG review, its power generation technologies were analysed. The information from 1,618 MGs around the world have been aggregated and studied. Also, two MG infrastructure model generators have been proposed (considering the infrastructure as the power generation technology and their rated power of an MG.). One of the models is based on the statistical data aggregated in tables and the other is based on ML techniques. The latter, which provides more particularised results, is able to generate the most typical MG infrastructure for a given location and segment of operation. Ideally, each of the DERs of a MG should be modelled, but, given the time constraints of a PhD, only the principal renewable generation technologies have been studied. Hence, ML models of photovoltaic (PV) systems plus a battery and wind energy conversion systems (WECS) have been proposed. Various ML models for PV systems were developed in two studies. First, an ML model for PV power estimation was performed using data from two real PV farms and validated using deterministic models from the literature. The ML algorithm was performed using neural networks and automatic strategies to clean the data. Neural network accuracy when trained and tested in the same location yields solid results which can be applied in performance ratio tools for PV power stations. In the second study, various mathematical models are proposed. This study provides several models for computing the annual optimum tilt angle for both fixed PV arrays and solar collectors. The optimum tilt algorithm proposed can be calculated in the absence of meteorological or software tools. To generate these models, data were collected from 2,551 sites across the world. A regression analysis with polynomial fitting, neural networks and decision trees was performed. Despite the better performance of the ML models, the ease of use of polynomial algorithms is recommended for those sites with no access to computational tools or meteorological data. The performances of the models were validated using previous research algorithms. Also, an ML algorithm was proposed to estimate the state of charge of a lithium-ion battery. The available capacity in a battery is an important feature when operating these kinds of systems. Given the complex behaviours of a battery, data-driven algorithms are able to capture the dynamic behaviours of a battery. Based on the data obtained in different experiments performed in a laboratory, an ensemble method, gradient boosting algorithm, was trained to model the state of charge of the battery. Even though the state of health of the storage system was below the theoretical life expectancy, the model was able to provide solid results. The model was validated with non-trained data. Finally, data-driven techniques were applied to model different elements of WECS. The first study provided two power coefficient algorithms, one based on polynomial fitting and another based on neural networks. To train the models, data from a corrected blade element momentum algorithm was used and three sets of data representing different wind turbine ranges, from 2 to 10 MW, were generated. Both models were validated with three datasets of real wind turbines and compared with the existing literature equations. Compared to previous equations, errors were reduced by at least 55% with the best numerical approximation from the literature. This type of reduction has a great impact for WECS dynamic and transient studies. The second study proposed for WECS develops three different ML models: one estimates the power of individual WECS, the second aggregates the data from all the WECS and estimates their power and the last one estimates the power of an entire wind farm. Given the stochastic and dynamic behaviours of the systems modelled, data pre-processing should be performed. Along with default cleaning techniques, a Student-t copula has been proposed so outliers can be automatically removed. Results show that the neural network algorithms’ performances for the three models can be improved without excessive manual intervention in the development process. Traditionally, electrical, physical and chemical models have been applied to mimic the behaviour of power systems. Now, with the power of computer and storage systems, a new era of more customised models has begun. It is time to review the existing models and provide better solutions by using ML techniques. In this thesis, only a few DERs have been modelled, but the results show that huge improvements can be made and future work in this subject should be done. The ML models proposed can be applied either as individual models for performance assessment of each DER or as a complementary tool to dynamic or static studies, unit commitment and planning software, among others.