Microgrid management with weather-based forecasting of energy generation, consumption and prices
The decade 2009-2019 was particularly intense in rhetoric about efforts to tackle the climate crisis, such as the 2015 United Nations Climate Change Conference, COP 21. However, the carbon dioxide emissions at the world scale increased constantly from 29.7 (GtCO2) in 2009 to 34.2 in 2019. The current gap between rhetoric and reality on emissions was and is still huge. The Intergovernmental Panel on Climate Change proposed different mitigation strategies to achieve the net emissions reductions that would be required to follow a pathway that limits global warming to 1.5°C with no or limited overshoot. There are still pathways to reach net-zero by 2050. Several reports propose detailed scenarios and strategies to achieve these targets. They remain narrow and highly challenging, requiring all stakeholders, governments, businesses, investors, and citizens to take action this year and every year after so that the goal does not slip out of reach. In most of these trajectories, electrification and an increased share of renewables are some of the key pillars. The transition towards a carbon-free society goes through an inevitable increase in the share of renewable generation in the energy mix and a drastic decrease in the total consumption of fossil fuels. In contrast to conventional power plants, renewable energy is subject to uncertainty. Most of the generation technologies based on renewable sources are non-dispatchable, and their production is stochastic and complex to predict in advance. A high share of renewables is challenging for power systems that have been designed and sized for dispatchable units. Therefore, this thesis studies the integration of renewables in power systems by investigating forecasting and decision-making tools. Since variable generation and electricity demand both fluctuate, they must be forecast ahead of time to inform real-time electricity scheduling and longer-term system planning. Better short-term forecasts enable system operators to reduce reliance on polluting standby plants and proactively manage increasing amounts of variable sources. Better long-term forecasts help system operators and investors to decide where and when to build variable plants. In this context, probabilistic forecasts, which aim at modeling the distribution of all possible future realizations, have become a vital tool to equip decision-makers, hopefully leading to better decisions in energy applications. When balancing electricity systems, system operators use scheduling and dispatch to determine how much power every controllable generator should produce. This process is slow and complex, governed by NP-hard optimization problems such as unit commitment and optimal power flow. Scheduling becomes even more complex as electricity systems include more storage, variable generators, and flexible demand. Thus, scheduling must improve significantly, allowing operators to rely on variable sources to manage systems. Therefore, stochastic or robust optimization strategies have been developed along with decomposition techniques to make the optimization problems tractable and efficient. These two challenges raise two central research questions studied in this thesis: (1) How to produce reliable probabilistic renewable generation forecasts, consumption, and electricity prices? (2) How to make decisions with uncertainty using probabilistic forecasts to improve scheduling? The thesis perimeter is the energy management of "small" systems such as microgrids at a residential scale on a day-ahead basis. The manuscript is divided into two main parts to propose directions to address both research questions: (1) a forecasting part; (2) a planning and control part. The forecasting part presents several techniques and strategies to produce and evaluate probabilistic forecasts. We provide the forecasting basics by introducing the different types of forecasts to characterize the behavior of stochastic variables, such as renewable generation, and the tools to assess the different types of forecasts. An example of forecast quality evaluation is given by assessing PV and electrical consumption point forecasts computed by standard deep-learning models such as recurrent neural networks. Then, the following Chapters investigate the quantile forecasts, scenarios, and density forecasts on several case studies. First, more advanced deep-learning models such as the encoder-decoder architecture produce PV quantile forecasts. Second, a density forecast-based approach computes probabilistic forecasts of imbalance prices on the Belgian case. Finally, a recent class of deep generative models, normalizing flows, generates renewable production and electrical consumption scenarios. Using an energy retailer case study, this technique is extensively compared to state-of-the-art generative models, the variational autoencoders and generative adversarial networks. The planning and control part proposes approaches and methodologies based on optimization for decision-making under uncertainty using probabilistic forecasts on several case studies. We introduce the basics of decision-making under uncertainty using optimization strategies: stochastic programming and robust optimization. Then, we investigate these strategies in several case studies in the following Chapters. First, we propose a value function-based approach to propagate information from operational planning to real-time optimization in a deterministic framework. Second, three Chapters focus on the energy management of a grid-connected renewable generation plant coupled with a battery energy storage device in the capacity firming market. This framework promotes renewable power generation facilities in small non-interconnected grids. The day-ahead planning of the system uses either a stochastic or a robust approach. Furthermore, a sizing methodology of the system is proposed. Finally, we consider the day-ahead market scheduling of an energy retailer to evaluate the forecast value of the deep learning generative models introduced in the forecasting part. (1) Forecasting techniques of the future. The development of new machine learning models that take advantage of the underlying physical process opens a new way of research. For instance, new forecasting techniques that take advantage of the power system characteristics, such as the graphical normalizing flows capable of learning the power network structure, could be applied to hierarchical forecasting. (2) Machine learning for optimization. Models that simplify optimization planning problems by learning a sub-optimal space. For instance, a deep learning model can partially learn the sizing space to provide a fast and efficient tool. A neural network can also emulate the behavior of a physics solver that solves electricity differential equations to compute electricity flow in power grids. Furthermore, such proxies could evaluate if a given operation planning decision would lead to acceptable trajectories where the reliability criterion is met in real-time. (3) Modelling and simulation of energy systems. New flexible and open-source optimization modeling tools are required to capture the growing complexity of such future energy systems. To this end, in the past few years, several open-source models for the strategic energy planning of urban and regional energy systems have been developed. EnergyScope TD and E4CLIM are two of them where we think it may be relevant to implement and test the forecasting techniques and scheduling strategies developed in this thesis. (4) Psychology and machine learning. Achieving sustainability goals requires as much the use of relevant technology as psychology. Therefore, one of the main challenges is not designing relevant technological tools but changing how we consume and behave in our society. Thus, machine learning and psychology could help to identify appropriate behaviors to reduce carbon footprint. Then, inform individuals, and provide constructive opportunities by modeling individual behavior.
