Since the introduction of SF6 in the 1950s, gas-insulated high voltage circuit breakers and Gas-insulated switchgear (GIS) have improved considerably, in particular concerning required drive energy for operation, compactness, and reliability. Nevertheless, high voltage insulation design has become increasingly challenging in recent years. Customer demands for reducing the physical footprint of HV equipment has led to an increase in operational field stress and therefore, much higher pressure on tolerances and increased susceptibility to defects. Dielectric design is based on how electrostatic fields are distributed and how much stress they can generate. In the course of conducting the intensive theoretical and experimental investigations on the dielectric design of insulation systems applied to high voltage power and pulsed power applications, it is becoming necessary to consider the influence of phenomena that have not been considered before. On top of that, as SF6 is one of the greenhouse gases listed in the Kyoto Protocol, SF6 usage regulations have been implemented in many industries. In Europe, SF6 is regulated under the F-Gas directive, which bans or restricts its use for several purposes. Several studies have indicated that CO2 is a viable alternative to SF6 for the transmission and distribution of electricity. In this context, this project aims to investigate a potential new phenomenon, the occurrence of secondary discharges caused by propagating streamer or leader discharges in HV gas insulation. Learning more about such phenomena can help us improve the dielectric design or perhaps explain the occurrence of breakdowns reported in high voltage equipment for which no obvious cause could be found. Specifically, CO2/O2 mixtures will be studied, and its results compared to SF6. Breakdown in gaseous insulation is caused by propagating discharges that are external in ambient air, e.g. on a bushing, or inside high-pressure insulation, e.g. across the surface of a GIS insulator. During the propagation of such discharges and flashovers (streamer or leader), transient electric field enhancement may occur, leading the local fields to exceed the inception fields at other locations. This can result in secondary discharges. Dielectric experiments using Image Intensifier and Photomultiplier tubes (PMT) have been conducted at three different pressure ranges; SF6 and CO2/O2 were selected as the insulation mediums. A negative polarity electric field is expected to trigger the secondary discharge. An analysis of images obtained from optical investigations and time lag records obtained from PMT signals suggested that secondary discharges can occur at low pressures (0.2MPa) in both SF6 and CO2/O2. At higher pressures (0.4 to 0.6MPa), no secondary discharges were detected. The reason for this is that, at higher pressures, the breakdown field is higher leading to a faster propagation of discharge across that gap. Hence, the formative time lag of the discharge is very short (some 100푛푠 or less) and this short time is not sufficient for secondary discharge inception.

The increasing adoption of renewable energy sources, particularly photovoltaic (PV) systems in residential sectors has raised important energy balancing challenges due to the intermittent nature of energy generation. To address these challenges and prioritize cost savings for residential consumers, this research investigates the integration of battery energy storage systems (BESS) and dynamic pricing strategies through an intelligent energy management system (EMS). Given the stochastic nature of PV generation, market prices, and load profile it is still challenging to achieve optimal control. Therefore Reinforcement Learning (RL)-based EMS is proposed in this research to make real-time optimal control decisions. RL is a machine learning approach where an agent learns to make decisions by interacting with an environment to maximize cumulative rewards. In this study, a deep deterministic policy gradient (DDPG) RL architecture is chosen due to its capability to handle continuous action spaces. In addition, deep learning-based models are employed to forecast uncontrollable load, PV generation and market prices for the integration into the EMS for which Bi-directional LSTM (Long Short Term Memory) was found to be the most accurate for all three uncertain variables. The DDPG algorithm is trained with data from a single household from the Lucerne region, Switzerland for 30 days and tested for a week. The results showed that compared to a deterministic rule-based approach the RL-based EMS increased cost savings for the end consumer by 14.2% but reduced the benefits for the grid operator to alleviate grid congestion quantified in terms of load factor, peak power consumption and ramping. Further work could be undertaken in testing the model on more extensive data and finding the best trade-off between customer and grid operator benefits.

Renewable energy sources, although they are quickly increasing their share in the energy mix, face a major barrier to more widespread adoption. Energy storage solutions overcome this hurdle, and lithium-ion batteries are at the forefront of this. The need for lithium-ion battery degradation studies arises due to the ever increasing use of these batteries in a wide variety of applications. There exist a large number of empirical and semi-empirical battery degradation models in literature, however their usage with irregular, real world profiles has not been explored. Moreover, understanding where different models can be used and comparing these models is also an area that has not been looked at. A real world power profile was created in MATLAB based on the WLTP drive cycle. Simultaneously, each model considered was verified against the experimental data for that same model to make sure model predicted degradation values matched the experimental observations. In three models, the verification procedure revealed errors and inaccuracies that were corrected. Implementing an irregular power profile posed two challenges, first, the stress factors of time and throughput required linearized versions of the model equations to be properly accounted for. Second, the identification of cycles in the irregular power profile is usually done using the rainflow counting algorithm, however, an alternative method is proposed in this study. The real world power profile was applied to each calendar and cyclic ageing model and it was noted that although the power profile was equalized for each cell based on the model in question, the results were widely differing. In some cases, it was possible to explain the differences but in some cases the differences were not easily explained. Apart from the comparison, each model was individually analyzed to understand how degradation phenomena and stress factors affect the accuracy and use case of a model. Finally, a toolbox was built in MATLAB to summarize the findings of this study, to help users understand where each model studied is likely to be accurate and useful. It was concluded that empirical and semi-empirical models are highly dependent on the testing conditions of the experimental data they are built on, and there are extremely limited scenarios in which these models are applicable outside these conditions. Moreover, accelerated testing conditions which are often used in the experimental phase usually cover different degradation phenomena than those which occur under regular use cases. With respect to the application of irregular (real-world) power profiles to these models, this study details a unique method used to obtain accurate predictions.

In this thesis, a new photovoltaic fault detection and classification method is proposed. It combines the generation of a synthetic photovoltaic training database and the use of a machine learning model to detect and classify faults in small-scale residential PV systems. The database was generated in Matlab, and the machine learning modeling was done with the scikit-learn library for Python. From the modeled PV systems, solely power yield is used as an indicator, combined with system age and meteorological conditions. Using these features, four types of machine learning models are used to detect malfunctioning PV systems and classify short-circuit faults and open-circuit faults. This thesis also shows the benefit of a synthetic PV training base as opposed to alternative methods, with increasing performance due to control of database balance. The result of this thesis is a method that can be used to construct a model for detection and classification of photovoltaic faults, specific to a single residential PV system. Malfunctioning systems can be detected with an accuracy of 80.4% using a random forest algorithm. For fault type classification, an F1-score of 0.759 was achieved, also using a random forest.

Distribution Network Operators (DNOs) are confronted with a significant challenge to expand the capacity of the electricity distribution grid to facilitate the energy transition. Grid expansion planning for the distribution grid is a complex problem with many constraints and objectives. Earlier research has focused on metaheuristics such as genetic algorithms to optimise grid expansion designs in terms of resolved congestions, redundancy and costs. However, this algorithm is not used in practice by its intended users; grid architects. This research first identifies why the algorithm is not yet used and proposes several ideas to improve user adoption. Then a technical implementation of sketch-based optimisation using a novel shape similarity measure is presented. We investigate its behaviour with a realistic case study. Furthermore, a qualitative user study is done to evaluate the potential impact of sketch-based optimisation on distribution grid expansion planning. We conclude that the novel shape similarity measure enables sketch-based optimisation of distribution grid expansion planning and that it has potential to accelerate electricity grid design processes. Alliander, the largest DNO of the Netherlands, will implement sketch-based optimisation in their distribution grid design application later this year, making it available to 20+ grid architects to accelerate their design process.

Distribution Network Operators (DNOs) are confronted with a significant challenge to expand the capacity of the electricity distribution grid to facilitate the energy transition. Grid expansion planning for the distribution grid is a complex problem with many constraints and objectives. Earlier research has focused on metaheuristics such as genetic algorithms to optimise grid expansion designs in terms of resolved congestions, redundancy and costs. However, this algorithm is not used in practice by its intended users; grid architects. This research first identifies why the algorithm is not yet used and proposes several ideas to improve user adoption. Then a technical implementation of sketch-based optimisation using a novel shape similarity measure is presented. We investigate its behaviour with a realistic case study. Furthermore, a qualitative user study is done to evaluate the potential impact of sketch-based optimisation on distribution grid expansion planning. We conclude that the novel shape similarity measure enables sketch-based optimisation of distribution grid expansion planning and that it has potential to accelerate electricity grid design processes. Alliander, the largest DNO of the Netherlands, will implement sketch-based optimisation in their distribution grid design application later this year, making it available to 20+ grid architects to accelerate their design process.

As the power system grows more complex and active, equivalent models have become a solution for modelling parts of the network that have limited observability or are confidential or too complex to simulate otherwise. In the past decade, this topic has also made its way to distribution networks because of its transition towards an active network, which was not the case before. The current grey- and black-box techniques for equivalent modelling create models that are fitted to the average dynamic response of the system or have limited applicability. Also, the existing methods lack extensive verification under different system conditions, and these works rarely focus on active distribution networks (ADNs) with multiple points of common coupling (PCCs). This thesis aims to develop an equivalent model that can estimate the dynamic response of an active distribution network with multiple PCCs and under diverse operating conditions and topological changes. The proposed approach is based on a graph-time convolutional neural network (GTCNN) that relates available PMU measurements inside the distribution network on a graph structure 𝒢GTCNN. The graph 𝒢GTCNN is obtained by taking a modified line graph of the graph representation of the power system and is expanded using the Cartesian product graph rule to include the temporal dependencies of nodes on their past values. The inputs of the equivalent model are the voltage magnitude |V| and angle θ at the PCC and the initial power injections P0 and Q at non-PCC nodes, while the model outputs the active P and reactive Q power at the PCC nodes. The GTCNN explicitly considers the initial power injections P0 and Q0 at non-PCC nodes to help the model learn how different operating conditions and topological changes impact the dynamic response. The DSO trains the equivalent model using simulation data or collected PMU measurements. The model is exchanged with the TSO every month, who can use the equivalent in co-simulation with their transmission network model to perform transient stability studies. The GTCNN-based equivalent model showed promising performance as an equivalent model for transient stability. The GTCNN was benchmarked against two state-of-the-art Long Short-Term Memory (LSTM)-based equivalent models and a hybrid GTCNN-LSTM model. The evaluation was performed on a real Dutch distribution network using three datasets, each focussing on a different system condition: different fault events, different operating conditions and different hidden topological changes. The GTCNN-based equivalent model had a mean-squared error (MSE) below 0.02 for each dataset, which means it can accurately reproduce the dynamics. This accuracy is comparable to the LSTM-based equivalent models, but the GTCNN could train 4x faster. The GTCNN also showed good generalisation performance, as its accuracy did not decrease on the validation and test sets. A study on scaling performance suggested that the MSE of the GTCNN-based equivalent model increases slower than that of the LSTM-based models while its training time increases faster. Therefore, the GTCNN-based equivalent model trains faster for smaller ADNs but will be more accurate with more measurement nodes. However, the proposed GTCNN has difficulty learning the response at different close-by PCC terminals if the dynamics are different. The developed GTCNN-based equivalent model can predict the dynamic response accurately under changing topologies and operating conditions at a similar performance level to existing LSTM-based approaches. However, its training time is much faster, which can result in a more accurate equivalent model by a more frequent model exchange between the DSO and TSO or a more extensive dataset being used to train the model. In future research, the GTCNN performance will be evaluated on a more comprehensive dataset containing all three system conditions to establish how much data is needed to train the equivalent model accurately. Also, the system frequency will be considered as an additional input. Moreover, its scaling performance will be evaluated more extensively and with a more efficient coding implementation. Furthermore, a heterogenous graph convolutional operator will be implemented to learn the connection per relational type (source node - edge type - target node). Finally, the co-simulation interface between the equivalent model and popular simulation tools will be explored.

Distribution System Operators (DSOs) are responsible preventing grid congestion, while accounting for growing demand and the intermittent nature of renewable energy resources. Incentive-based demand response programs promise real-time flexibility to relieve grid congestion. To include residential consumers in these programs, aggregators can financially incentivize participants to reduce their energy demand and make aggregated energy reduction available to DSOs. A key challenge for aggregators is to coordinate heterogeneous preferences from multiple participants while preserving their privacy. This thesis proposes MARL-iDR: a decentralized Multi-Agent Reinforcement Learning approach to an incentive-based demand response program. The approach respects participants' privacy and preferences and makes decisions in real-time when deployed. The aggregator and each participant are controlled by Deep Reinforcement Learning agents that learn to maximize their reward. The aggregator agent learns a policy that dispatches suitable incentives to participants based on total energy demand and a target reduction, while minimizing financial costs. The participant agent learns to respond to these incentives by reducing consumption to a fraction of the original demand. The participant agents curtail or shift requested household appliances based on the selected consumption reduction using a novel Disjunctively Constrained Knapsack Problem optimization, while minimizing residents' dissatisfaction. A case study with real-world electricity data from 25 households demonstrates the capability to induce demand-side flexibility. The approach is compared to the case without demand response and to a centralized myopic baseline approach. A 9% reduction of the Peak-to-Average ratio (PAR) was achieved compared to the original PAR (no demand response).

Solar energy is an abundant, scalable, and clean source of energy. With an exponential drop in prices of PV modules, more and more rooftop photovoltaic (PV) systems are being installed worldwide. Since these small-scale PV systems do not use expensive sensors, it is difficult to detect malfunctions for these systems. This could lead to lower energy generation along with financial losses for the owners. Thus, a method for PV yield monitoring is developed for early and remote fault detection. This method does not use the conventional analytical approach as it depends on inaccurately extrapolated weather data. Instead, the proposed method uses data from similar or neighbouring or peer PV systems for estimating the expected energy generation. By comparing the expected energy generation with actual energy generation, a faulty system can be flagged. In this project, information from about 12000 PV systems is used, which includes system design information such as location, number of panels, panel orientation, etc. along with the historical daily energy generation for periods ranging from two months to up to seven years per system. In this thesis, a machine learning model was developed for predicting energy yields, which uses a Genetic Algorithm (GA) for optimization. This model splits the available data into system design, system location and system yield data. Thus, the model uses these as criteria for finding PV systems similar to the monitored system. Once good peer systems are located, system yield data of those systems are used for estimating the expected energy yields of the monitored system. The three criteria used by the model do not have equal influence on finding good peers, thus, the model had to be trained or optimization was done using the training data. Post optimization, the relative influence of system design: system yield: system location was found to be 0.125:0.875:0 with on average 16 good peers needed for accurate predictions. The proposed model has a mean normalized RMSE of 0.057 and about 95% of the systems tested had an R2 score higher than 0.85. The existing commercial software at Solar Monkey has a mean normalized RMSE of 0.082 and about 83% of the systems tested had an R2 score higher than 0.85. The predicted energy generation calculated by the proposed model is compared with the actual energy generation to detect any malfunctions that may have occurred in the monitored system. Thus, 120 randomly chosen PV systems were analysed for faults. Based on this, a semi-automatic categorization framework was created with the proposed model as one of the criteria to detect common faults in the system such as missing data, under-performance, over-performance and false positives. Using the categorization framework, certain PV systems were found as interesting examples for under-performance with broken panels or string, over-performance with system size change and false positives. The model is especially useful for separating system design mismatch from actual system malfunctions. With the framework, it was shown how the proposed peer-to-peer model can be used for fault detection along with certain other models.

The transition to green energy is reshaping the energy landscape, marked by increased integration of renewable energy sources, distributed resources, and the electrification of other energy sectors. These changes challenge grid security, particularly regarding the N-1 security criterion, a crucial factor in preventing blackouts. Furthermore, climate change is contributing to the growing frequency of extreme weather events, which constitute the second major cause of blackouts. As grid complexity keeps on increasing, the need for N-k security, where k lines fail simultaneously, and increased resilience against extreme weather events is becoming increasingly evident. This necessitates studying the security constrained optimal power flow (SCOPF) problem considering multiple line outages (N-k). Current methods exhibit poor scalability as k increases. In response to the challenge of limited scalability, this thesis proposes a constraint-driven machine learning approach to approximate N-k SCOPFs. The proposed approach relies on the linearized direct current optimal power flow. The approach utilizes a neural network to map power system loads to generator setpoints. A feasibility restoration layer is employed to restore base case infeasible predictions. By incorporating line outage distribution factors (LODFs), all post-contingency flows are computed. The loss function utilized to train the neural network draws inspiration from the penalty function method. Lastly, a copula analysis computes joint outage probabilities for k extgreater 1 enabling a probabilistic security assessment. The first academic contribution of this thesis is the development of a constraint-driven approach to approximate N-k SCOPFs considering all contingencies using LODFs. The second academic contribution is the formulation of a N-k risk based security criterion, providing an alternative to the current deterministic N-1 security criterion. The approach shows promise in its ability to scale effectively to N-k contingencies. Using LODFs, the approach effectively computes all post-contingency flows for up to k = 3. Moreover, case studies show the constraint-driven approach's effectiveness in identifying violating post-contingency cases, with up to 173$ imes$ speedups and close to optimal dispatch costs. However, the consideration of N-k contingencies holds combinatorial complexity, and more efficient methods need to be developed for the computation and storage of all LODFs, and for the computation of all post-contingency flows. Additionally, the proposed constraint-driven approach can not enforce any post-contingency constraints, necessitating post-contingency feasibility checks when security against specific contingencies is required. Next, by incorporating probabilities, the approach shows promise in improving power systems security and resilience, but further research is necessary. In this thesis, only line outages are considered. In the future, the approach could be modified to additionally account for other equipment outages (e.g. generator outages). Furthermore, future research could investigate the adoption of this approach in corrective control settings, where it is employed in the restorative phase of a contingency event. Another suggestion is centered around the incorporation of graph neural networks in the proposed approach, which could provide a more scalable alternative to fully connected linear neural networks. Furthermore, more scalable methodologies could be explored to construct the matrix containing all LODFs, and a more scalable methodology for computing all post-contingency flows could be developed. Finally, future work could investigate how to utilize the proposed approach under varying conditions like network topology changes or changing outage probabilities.

The growth of renewable energy technologies is leading to energy systems that are more reliant than ever on renewables such as Wind and Photovoltaic (PV) power. Despite their benefits in terms of sustainability, their ubiquity poses challenges in maintaining grid stability given their intermittency, emphasising the prediction of power fluctuations. Physical models and statistical approaches, especially for nowcasting (forecasting for 0-6 hours in the future), have been superseded by Machine Learning (ML) methods in terms of forecast accuracy (below 3% Root Mean Squared Error (RMSE)). Within ML, Artificial Neural Network (ANN) methods seem to perform particularly well for nowcasting. This project focuses on predicting solar and wind meteorology with that level of accuracy, and on how to best use the prediction to minimize the cost of maintaining a balanced energy system, i.e. one where power consumption matches production at any moment. Producing accurate power predictions based on Multi-Modal (MM) data and the extent to which prediction accuracy reduces system cost are challenges to be addressed in this thesis. MM and End-to-End (E2E) training (with the system cost as the task of an ANN based algorithm) are investigated to this end. MM learning involves handling information from multiple types of input (audio and visual, for example) for performing a ML task such as regression or classification. It is of interest for this project because it has been shown to outperform other NN approaches in predicting sudden changes in solar irradiance. E2E learning entails an algorithm design which predicts the end goal of a ML process directly from the inputs. This is pursued because it addresses the true task (cost minimization) of system operators as the focus of the ML algorithm. The proposed method consists of a NN architecture that learns to fuse features from MM data (sky imagery and meteorological sensor data) at intermediate layers of the network in order to predict PV or Wind generation. This prediction is then used as an input to an Optimal Power Flow (OPF) problem (which seeks to minimize generation costs in a power system, considering power balance and transmission network constraints to ensure the twin goals of economic and secure system operation). The proposed model is trained E2E, therefore it is informed by the minimized cost solved by the optimization, rather than the intermediate power prediction (as conventional approaches would involve). In an IEEE 6-bus system with PV generation, a sequential training baseline results in costs 10% higher than a perfect forecast, while our proposed MM4-E2E approach achieves costs only 7% higher, a significant improvement. The intermediate prediction of PV power by MM4-E2E is also improved, with 18% lower RMSE by the proposed model compared to the baseline, explained by the enhancement of one modality by the other through MM learning. In a power system with two renewable sources, costs are reduced through the proposed model compared to a conventional approach (4% excess cost compared to 7%, measured against a perfect forecast), but power prediction accuracy is worse, sue to convergence to local minima.

Short-term solar forecasting is crucial for large scale implementation of solar energy and plays an important role in grid balancing, energy trading, and power plant operation. Cloud movement is the main source of unpredictability within solar forecasting and can be recorded using All-Sky Imagers. Conventional cloud modelling methods using image analysis techniques are unable to extract the spatial configuration and the temporal dynamics of clouds, resulting in poor predictions of the interaction with solar radiation. The goal of this study is to create a deep learning model for short-term irradiance forecasting between 0 and 21 minutes into the future using all sky images combined with auxiliary data. The model performance was assessed by comparing the deep learning model with the persistence model and showed that the deep learning model outperforms the persistence model with 24.8%. A sensitivity analysis to data usage is performed showing that besides using more data, also the variation of using multiple years of data results in better performance. Furthermore, the sensitivity of the model to input variables is assessed, showing that using the clear sky irradiance as input improves model performance with 16% and that meteorological data does not improve performance. Additionally, the model performance was evaluated during different sky conditions showing that the deep learning model outperforms the persistence model for all sky conditions, except overcast conditions. An example of the model behavior is extensively described, showing that the deep learning model tends to predict the trend of the irradiance fluctuations rather than the actual fluctuations. Next to that is in this study shown that the current deep learning model occasional miss important weather events, like obscuration of the Sun, resulting in large irradiance prediction errors. A pathway for future improvements for deep learning models to forecast the short-term irradiance is provided.

Whereas in the past, Distribution Systems played a passive role in connecting customers to electricity, Distribution System Operators (DSOs) will have to take in the future a more active role in monitoring and regulating the network to deal with the new behaviors and dynamics of the system brought by the energy transition. State Estimation, a task traditionally reserved for Transmission System Operators (TSOs), is, therefore, a needed tool for DSOs to properly monitor the distribution grid in the future. However, the implementation of Distribution System State Estimation (DSSE) faces several challenges. The distribution system lacks observability to get satisfying estimation accuracy, the denser network increases the complexity of the estimation process, and the lack of labeled data makes training Machine Learning alternatives difficult. To tackle these issues, we propose the Deep Statistical Solver for Distribution System State Estimation (DSS2), a Deep Learning model based on the Graph Neural Network (GNN) architecture and the Physic-Informed Machine Learning (PIML) framework. The DSS2 model is based on the Deep Statistical Solver (DSS) framework, which seamlessly models power systems into GNN using Hyper-Heterogeneous Multi Graphs (H2MG), and emphasizes semi-supervised learning by learning to optimize, using optimization problem as a loss function. This thesis extends the DSS framework to the DSSE problem, using the traditional State Estimation algorithm as an optimization problem to learn, and incorporating the power flow equations in the loss function. This model is trained through a semi-supervised approach to learn the physics of the problem and alleviate the need for labels and uses the Deep Learning tools to improve accuracy and robustness in the DSSE task. Case studies on 14-bus, 70-bus, and 179-bus networks show promising results, with the model outperforming the traditional WLS algorithm while showing better robustness. The model also competed in performance against supervised models and showed to be more suitable for the semi-supervised learning approach than simpler GNN architectures.

This work seeks to resolve an outstanding problem in the use of reinforcement-learning methods for the simulation of economically-rational agents. We discuss the problem of non-stationarity, and how this subsequently limits market simulation capabilities. After explicating and isolating the source of the problem for a day-ahead electricity market, we demonstrate the application of methods which resolve this problem in simple test-cases, and prove conditions under which similar methods will work in general. Subsequently, we illustrate how these techniques can be used to solve a restricted market-design problem, in the process introducing a framework for discussing adversarial market-design for electricity markets in general. It is hoped that, insofar as they provide a new feedback-loop for market-design, these results will facilitate the design of more complex electricity-markets suitable for the energy transition.

To keep pace with increasing renewable energy penetration and consequent increase in inverter-based resources in the power grid, it is pertinent for present-day research to address the resulting drop in system inertia levels and its impact on frequency stability. With decreasing levels of inherent rotational inertia present in the system, any sudden disturbance causing an energy imbalance in the grid could lead to more drastic excursions of system frequency than those experienced hitherto. To ensure the resilience of the grid in such scenarios, advanced and competent frequency stability assessment and control methods are required. This thesis presents Neural Ordinary Differential Equations (NODE), a recently introduced family of neural networks, as an effective tool to achieve fast, real time estimates of the expected frequency response trajectory during an energy imbalance event. Since high-impact frequency instability events are sparse in reality, both real-world grid data and synthetically generated data corresponding to different inertial conditions are used to train predictive NODE models. Firstly, NODE is adapted to frequency prediction applications through relevant data processing steps, and modification of network parameters and algorithmic aspects pertaining to the predictive model definition. Secondly, patterns corresponding to specific sections of the frequency response curve are used to selectively train NODE models. Pattern-specific training methods exhibit better prediction performance when the NODE model encounters frequency behaviour similar to the one it initially trained on. Thirdly, a pre-training approach to cut short on the real-time training time required by NODE models to achieve desired levels of prediction performance is presented. Fast estimates of critical frequency stability parameters like nadir could act as potential triggers for early stability control actions to achieve a more controlled frequency response. Application of predictive NODE models for different frequency scenarios are presented using three test-cases: normal operating scenario, restoration post-system split scenario and synthetically generated high-impact frequency disturbance scenarios. Model tuning and training methods specific to each test-case are described, and prediction results are evaluated with relevant performance metrics. Finally, a comparison is made between the implementation of NODE among different test-cases and real-world implications of the frequency prediction outcomes from the test-cases are further discussed.

Renewable energy generation projects are often measured by their peak capacity. A wind farm rated at 25 MW will generate 25 MW of power under the right circumstances. This peak capacity is reached very little in practice. However, these generators are forced to purchase grid operation infrastructure that can handle this peak generation event. The high voltage grid connections are expensive and increasingly difficult to receive permits for. This work presents a solution in which the high voltage grid connection is undersized in comparison to the renewable energy generator. A battery energy storage system is installed in the local grid to solve the issue of excess energy generation (congestion). A simulation of the local network has been built that models a battery energy storage system (BESS), the network and uses data from a solar park. A case study in which a 19 MW solar park is connected to the high voltage grid with a transformer of only 14 MW as well as a 14 MW | 30 MWh BESS on the network is investigated in the rest of the work. Furthermore a BESS control strategy for price arbitrage on the TenneT imbalance market is presented and encoded such that it can be optimised. Four heuristics are presented that time and size the congestion issue in a manner the control strategy of the BESS can prepare for and solve congestion when necessary. These heuristics are tested against strategies optimized for revenue maximisation through price arbitrage. While the most aggressive strategies did not solve all the congestion events in these simulations, we found that the heuristic that takes the average generation of the solar park into account performs the best while remaining appropriately conservative. A basic evolutionary algorithm is presented that optimizes a BESS control strategy for price arbitrage when the BESS is not needed on the local network to solve congestion. Although the strategies earn ~33% less revenue due to the congestion related limitations, the optimisation surrounding congestion does improve revenue by 2.58%. The results presented in this work suggest that this setup of a local grid can be economically viable and that the BESS can solve the congestion issue when steered with an appropriate control strategy. We hope to inspire parties that battery energy storage systems can earn substantial revenue aside from solving issues on a (local) grid.