Conference Agenda

As energy systems around the world become increasingly dependent on variable renewable energy sources, optimal spatial and temporal balancing over different distances and time scales by means of networks and storage has become a central question of interest. Long-duration storage (LDS), based on green hydrogen or its derivatives, is envisioned to take a pivotal role in solving it. Formulated as deterministic linear programs, most energy system planning models determine optimal LDS capacities based on one or few weather years, assuming perfect foresight in the dispatch stage leading to potentially highly sub-optimal capacity choices. Recasting the planning problem as a multi-stage stochastic program with a planning stage followed by several operational stages, we employ a stochastic dual dynamic programming (SDDP) algorithm to capture inherent weather uncertainty and to abandon the perfect foresight assumption.

Relevant model sizes require computational refinements to the original SDDP algorithm to remain tractable. We adopt and compare several extensions suggested in recent literature contributions and test their efficacy for the energy system planning problem at hand. These extensions include stabilisation and warm starts, quasi-Monte-Carlo sampling schemes, inexact oracles and Chebyshev cut generation.

The share of renewable energy generation capacities will rise substantially in the upcoming decades. Their integration requires expansion of power grids and flexibility options, such as storage solutions. Current research in the field of energy system modelling therefore often aims at optimizing these technologies jointly, resulting in complex optimization problems to be solved. Their structure reveals decomposability on the temporal as well as on the spatial scale, which allows the application of decomposition techniques to keep them computationally tractable. Investment decisions and storage levels represent linking variables on the temporal scale and power exchanges between regions can be defined as linking variables on the spatial scale. Suitable decomposition techniques for dealing with linking variables are Benders decomposition and Consensus ADMM. Within this work we analyze which of these techniques is most favorable for solving large-scale multi-regional energy system models. Furthermore, we analyze which decomposition of the original problem, temporal, spatial or combined, can be solved most efficiently by which algorithm. We deliver a case study for the comparison of different decomposition strategies for a specific problem type and contribute to the open question of which decompositions are most suitable for which type of optimization problems.

Energy system optimization models (ESOMs) are useful tools to analyze the decarbonization of our current energy system. However, the underlying input data is subject to uncertainties. Although these uncertainties can substantially impact the structure of the energy system, they are often not taken into account. By applying stochastic programming (SP) the uncertainties can be considered within a single optimization run to achieve risk hedging.

Even without considering uncertainties, ESOMs can become exceedingly large when a high spatial and technological granularity is needed, e.g. for sector coupling. Therefore, acceleration techniques are required to keep the models solvable, especially when taking SP into account. Benders Decomposition (BD) is a method that is typically applied to SP models. Since the stochastic scenarios can be independently optimized within the subproblems, they can be easily parallelized along the stochastic scenario dimension. However, the number of scenarios is rather small in comparison to the number of time steps that are considered in hourly resolved ESOMs.

To further exploit the parallelization potential of stochastic ESOMs, we apply an additional decomposition along the time dimension to two different methods. First, we apply time-splitting to BD in combination with MPI. Second, the performance is compared to the parallel high-performance computing solver PIPS-IPM++. The solver, mainly applied on temporally decomposed ESOMs, has recently been extended to also incorporate stochastic optimization, which enables an additional decomposition along the stochastic scenario dimension.

Climate change mitigation requires supplying the power, heating, and transport sectors with renewable energy sources such as wind, solar, and hydro. Renewable supply and energy demand vary substantially across seasons or years, change with the climate, and cannot be predicted in the long-term. As a result, periods with high residual demand ranging from Dunkelflauten to multi-year energy draughts threaten energy security.

This work investigates how to secure renewable energy systems against climate uncertainty. We introduce a stochastic programming formulation that remains computationally tractable but can consider an extensive sample of weather conditions and limit foresight. In addition, we deploy a stabilized Benders algorithm to leverage distributed computing when solving the resulting stochastic optimization problem.

Building on this method, we analyse energy security in an interconnected European energy system characterized by fluctuating renewable supply and electrification. The analysis considers seasonal and intra-annual storage from hydro reservoirs or synthetic fuels, importing electricity or synthetic fuels, and renewable overcapacity or load shedding as security options.

Samples for uncertain and weather-sensitive energy system inputs build on ERA5 reanalysis data from 1982 to 2016. Thanks to the introduced formulation, capacity planning must not consider the entire dataset but relies on a subset to reduce problem size. A subsequent Monte-Carlo analysis with an operational rolling horizon model estimates the expected unserved energy for the computed capacity setups using the entire dataset.

Preliminary results show that energy security is critical to include in future energy planning but, at the same time, manageable when addressed correctly.