Decarbonization of Power Systems, Multi-Stage Decision-Making with Policy and Technology Uncertainty

Abstract

There is widespread agreement that "deep decarbonization" of the power sector, i.e., reduction of CO2 emissions to near or below zero, will be pivotal to climate change mitigation efforts. Nevertheless, given multi-decadal time horizons, planning for decarbonization must contend with uncertainty regarding technologies costs, new technologies characteristics and availability, and policies and incentives for reducing CO2 emissions in the multi-year adaptation process. At the same time, increasing penetration of variable renewable energy, the availability of energy storage technologies and the active participation of demand in electricity systems requires the appropriate consideration of temporal and spatial resolution to properly account for the cost and value of different resources at the system level.

New approaches are required to determine a the multi-year capacity expansion problem with hourly detail is computationally intractable using current methods and computational resources due to the increased number (millions) of variable and constraints that are involved in the problem. Our framework turns the problem into a computationally tractable one. This is accomplished by means of three decomposition methods at different levels and the integration of such methods into a single computational framework (FLIP).

Stochastic Dual Dynamic Programming is used to break down the problem at the year level iteratively passing information forwards and backwards across different years; Benders Partitioning is used to separate each yearly problem into a master investment problem and an operational problem passing information upwards and downwards between the two levels; and Dantzig-Wolfe decomposition is used to separate the year-long operational problem into a simplified operational problem and many operational sub-problems (e.g. weekly) passing information across problems to coordinate the year coupling constraints (e.g. CO2 policy) iteratively to find the optimal operation. This integrated framework requires solving orders of magnitude greater numbers of problems that are orders of magnitude smaller in size and complexity (number of variables and constraints down to thousands or hundreds from millions).

At the same time, the framework allows for parallelization at different levels of the problem, allowing the user to harness high performance computing resources with greater flexibility. The framework is implemented using the Julia general purpose programming language and its mathematical programming extension JuMP. value to the power systems of the 21st century. Conventional cost-based metrics (e.g., LCOE) are incapable of accounting for the indirect system costs associated with intermittent electricity generation, in addition to environmental and security constraints. Moreover, as recent research has shown, commonly used abstraction methods (sample hours, days or weeks selection methods) can provide inaccurate results by undervaluing some resources and overvaluing others.

Hence, there is a need to account for greater detail at the operational level while also accounting for multidecadal scenarios and imperfect information regarding costs, technologies and policies, all within a framework that is able to capture the value-cost trade-off dynamics of electricity resources. This work develops a methodology to properly account for the value-cost dynamics at the system level for decarbonization of power systems. Using this methodology, it then explores two key questions for policy and decision makers. First, we study the role of firm low carbon resources for deep decarbonization of power generation. We find that availability of firm low-carbon technologies -- including nuclear, natural gas with carbon capture and sequestration, and bioenergy --

reduces electricity costs by 10-62% across fully decarbonized cases. Then, we study the role of long duration energy storage (LDES) technologies for deep decarbonization. We find that the total system of LDES must fall below 40 [$/kWh] for LDES technologies to reduce system cost by more than 10%, even in our best case scenario. Finally, we expand our methodology into a multi-year capacity expansion planning framework for power systems that is able to solve for the optimal investment strategy/pathway with respect to future policies such as CO2 limits and/or renewable energy mandates while accounting for detailed operation at an hourly resolution over a full year as well as highlevel uncertainty (e.g. policy commitment, technology availability, etc).

In its original form the multi-year capacity expansion problem with hourly detail is computationally intractable using current methods and computational resources due to the increased number (millions) of variable and constraints that are involved in the problem. Our framework turns the problem into a computationally tractable one. This is accomplished by means of three decomposition methods at different levels and the integration of such methods into a single computational framework (FLIP).

Stochastic Dual Dynamic Programming is used to break down the problem at the year level iteratively passing information forwards and backwards across different years; Benders Partitioning is used to separate each yearly problem into a master investment problem and an operational problem passing information upwards and downwards between the two levels; and Dantzig-Wolfe decomposition is used to separate the year-long operational problem into a simplified operational problem and many operational sub-problems (e.g. weekly) passing information across problems to coordinate the year coupling constraints (e.g. CO2 policy) iteratively to find the optimal operation. This integrated framework requires solving orders of magnitude greater numbers of problems that are orders of magnitude smaller in size and complexity (number of variables and constraints down to thousands or hundreds from millions).

At the same time, the framework allows for parallelization at different levels of the problem, allowing the user to harness high performance computing resources with greater flexibility. The framework is implemented using the Julia general purpose programming language and its mathematical programming extension JuMP.