The decentralisation of energy supply, largely driven by renewable energy sources, has led to increased volatility and imbalances in the power grid. To maintain grid stability, flexible balancing mechanisms are required. Ancillary services offer this flexibility by either injecting power (upward regulation) or absorbing it (downward regulation). Among these services, the Automatic Frequency Restoration Reserve (aFRR) plays a crucial role due to its large regulating capacity and direct impact on imbalance prices through its activation costs. As market volatility increases, short-term trading in intra-day and imbalance markets becomes more important, increasing the need for accurate forecasting. While day-ahead and intra-day market forecasting is well-established in Electricity Price Forecasting (EPF), there is limited research on imbalance markets, especially the Dutch aFRR market, highlighting the novelty of this study. The Dutch market uses a dual pricing system with separate bid ladders for upward and downward regulation.

This research focuses on forecasting aFRR bid ladders in the Dutch electricity market and explores their application in short-term trading strategies. Forecasts are generated three hours ahead of delivery, aligning with decision points for intra-day trading, aFRR participation, or opting for no action. The study includes a detailed market analysis, a review of forecasting methods, the development of a machine learning framework tailored to the aFRR market, evaluation of forecast performance, and the practical use of these forecasts in trading scenarios.

A structured machine learning pipeline is designed, encompassing data pre-processing, transformation, model selection, prediction, and evaluation. During the transformation phase, data is scaled and then reduced in dimensionality using Principal Component Analysis (PCA) to retain key variance. The resulting components are used as inputs for predictive models, including LASSO, XGBoost, and LSTM, which are benchmarked against preliminary bid ladders published three hours before delivery. Forecast accuracy is evaluated using point metrics (sMAPE), interval metrics (PICP and PINAW), and a novel self-developed metric called the Largest Knick Volume (LKV), which captures accuracy at key inflection points in the bid ladders that are most relevant for short-term trading.

The findings are twofold. First, model evaluations both at the PCA level and on the reconstructed bid ladders indicate that all models can track general market trends but do not outperform the benchmark. Diebold-Mariano tests confirm that benchmark performance is superior at the PCA level. After reconstructing the bid ladders, benchmark sMAPE scores are 7% for upward and 8% for downward regulation, outperforming LASSO (7%/10%), XGBoost (8%/11%), and LSTM (7%/11%). Second, the forecasts are integrated into Battery Energy Storage System (BESS) intra-day trading strategies, including one based solely on intra-day prices and three incorporating different aFRR bid ladder positions: gas turbine marginal cost, intra-day price with a premium, and LKV-based positioning. Integrating aFRR forecasts improves trading performance, with some high volume-price strategies boosting revenue by up to 18%.

Although the developed forecasting methods do not outperform the benchmark, the benchmark itself already captures much of the relevant market information available three hours ahead, limiting the scope for additional forecasting improvements. This limitation is attributed to the weak correlation between market fundamentals and bid outcomes, as well as the strong influence of individual actors in the relatively small Dutch aFRR market. Nonetheless, the study shows that using existing market data and participating across multiple markets can improve profitability. The results support the strategic value of data-driven, forecast-informed bidding for enhancing the economic performance of battery storage systems.

Under the increasing electrification of end uses in the energy transition towards more renewable integration, the electricity price keeps gaining importance on every scale from individual well-being to the competitiveness of an economy. Though scarce in the scientific literature, Long-Term Electricity Price Projection (LEPP) has great potentials in decision-making and planning, as well as complementing the long-term energy scenarios. This study takes features from the Dutch, Spanish and Danish data in five years (2015-2019) to train deep neural networks in the conditional Wasserstein Generative Adversarial Nets with Gradient Penalty (cWGAN-GP) framework, in order to project Day-Ahead Market (DAM) price series under Dutch 2050 energy scenarios.

The LEPP to 2050 is made possible by normalising the selected markets. As a result, the conditions unprecedented in Dutch data are covered in the normalised and combined data set. Generally, under scenarios with high proportions of hydrogen power in the energy portfolio, the cWGAN-GP model projects that DAM price series would have slightly lower mean and daily standard deviation than the 2019 level. Whereas much lower mean and daily standard deviation are projected when natural gas is still the fuel of the most frequent final generating technology. To explore the possible application of the projector model, the German DAM prices series in 2019 have been projected and evaluated, and the projections under Dutch 2050 energy scenarios have been used in calculating the generic profit potential of energy storage.

Five findings can be summarised from the main results. Firstly, from a literature survey and importance analyses, seven features are shown relevant to the DAM price in the combined data set, namely month of the year, day of the week, total hourly load forecast, national daily mean temperature, fuel cost of the most frequent final generating technology, hourly renewable power generation forecast and total installed renewable power capacity. Secondly, it has been found that two of the four proposed market state normalisation solutions, the Renewable Scarcity Factor (RSF) and the Renewable-Load Ratio (RLR) help the cWGAN-GP model strike a balance between price value distribution and hourly inter-dependencies. Thirdly, in this LEPP study, the cWGAN-GP model performs better than the Conditional Variational Auto-Encoder (CVAE) and multivariate Gaussian distribution (mGaus) models. Compared with the two alternatives, the cWGAN-GP model produces samples in better quality while remaining sensitive to temporal conditions. Fourthly, projections by the cWGAN-GP model are more realistic than those made by the Energy Transition Model (ETM), with price values varying continuously in smooth boundaries. Finally, the fuel cost of the most frequent final generating technology is found critical to LEPP. The annual mean and daily standard deviation of the DAM price series are expected to rise significantly when natural gas is mostly replaced by hydrogen power in the national energy portfolio.

The production and consumption of electricity need to be balanced at all times. Due to the ever-growing shift towards renewable energy generation, this poses an increasingly difficult challenge. Currently, supply is regulated to maintain balance. However, there is potential to improve reliability and save costs by shifting the balancing to the demand side, known as demand response. The flexibility of water systems can play a role in this, thereby benefiting from cheaper price fluctuations and reducing operating costs.

This research investigates the IJmuiden pumping station, which drains water from the Noordzeekanaal-Amsterdam-Rijnkanaal system into The North Sea. The primary focus of the control of this system is ensuring safe water levels as it runs through areas of high economic value. The flexibility of the range of safe water levels allows costs to be minimized by selecting favourable moments to consume electricity. This simultaneously contributes to the stability of the electrical grid. This research explores the potential for a Reinforcement Learning controller for such an optimization problem, as there are some drawbacks to the Model Predictive Control methods that are currently widely used. The research objective is formulated as follows:

To optimize the control of the IJmuiden pumping station using Reinforcement Learning while complying with local water level restrictions and compare it to the state-of-the-art Model Predictive Control methods in terms of constraint violation, energy costs, and computational speed.

The Reinforcement Learning controller will use a deep Q-learning algorithm that chooses the most cost efficient control in IJmuiden while respecting the water level restrictions. To do so, the model makes decisions based on electricity prices and details about the state of the water system for the current time step as well as a forecast of 48 hours ahead. This data is provided as an input to the model.

The inputs of the model consist of historical data, meaning that the associated uncertainties are not included. The water system that the model can interact with is represented by a linear reservoir model. Therefore, the water system is influenced dynamically by the actions taken by the model. The possible actions are determined by the state of the water system.

The trained model was tested on 2 years of unseen data (data that was not used during training). Using the same test data, control plans were generated using Model Predictive Control. The Reinforcement Learning model was very successful in ensuring safe water levels. However, this did result in approximately 50\% higher energy costs. The use of the gate was close to optimal but the pumping was not clearly correlated with favourable prices and power consumption. The trained model was robust, with consistently accurate results with regards to respecting the water level constraints.

The most significant difference with the Model Predictive Control was the computation time. The Reinforcement Learning model was able to create a control plan approximately 300 times faster. This opens doors for further development of the model and increased complexity. A more accurate model of the water system can be used to take into account temporal and spatial effects and individually representing the six pumps in IJmuiden.

There are still many steps before such a model can be used for operational control, but the method has potential for such an application. Many aspects of the model can be improved as well as making adjustments to increase the usability for control operators.

This research provides benchmark accuracies for forecasting of an aggregated price of the Dutch intraday market. While point forecasts in a single-step-ahead horizon for that unresearched market provide novel insights already, the scope of this research also includes interval forecasts in a multi-step-ahead horizon. A forecasting procedure is established that organizes several stages of in-sample and out-of-sample testing so that the number of arbitrary choices regarding features and hyperparameters is kept as low as possible. It is concluded on the basis of accuracies attained by naive, regression, and artificial neural network models that the machine learning models that are capable to incorporate linear and nonlinear relationships are able to infer to a varying degree what drives intraday from day-ahead prices. Furthermore, it is addressed whether superiority in terms of accuracy coincides with what is deemed as superior in practice. A simulation of a generic system, which consists of a battery and a wind turbine located in the Netherlands, smartly dispatches stored energy according to a schedule optimized with model predictive control based on point forecasts of intraday price. It is concluded that, in general, slightly higher profits are obtained with more accurate point forecasts and that different point forecasts lead to very different dispatch schedules that vary more than 10% in terms of dispatch frequency.