REopt Transition to Open-Source Solvers

[kkrah]

The REopt Web Tool and API have transitioned to open-source optimization solvers (specifically, HiGHS, Cbc, and SCIP), which may lead to longer solve times and timeout errors. While we plan to continue tuning integration of these solvers to improve solve times, please expect slower-than-usual solve times until further notice.

Please see the guidance below for additional information about this switch and suggestions to improve solve times with these open-source solvers by simplifying the problem and/or adjusting solver settings.

Simplify the problem: Certain capabilities in REopt are more challenging for open-source solvers. Avoiding capabilities with complex formulations (e.g., net metering, resilience analyses, tiered utility rates, and minimum turndown on generator and CHP) and constraining the solution space (minimum and maximum system sizes being considered) can help speed up solve times.

• Constrain solution space: By default, REopt evaluates candidate technologies across a wide range of system sizes. If you have a sense of likely system sizing for your site, you could enter a maximum and/or minimum system size for each candidate technology. (For instance, if your site’s peak demand is 10 kW, it probably wouldn’t make sense to use a 100 kW PV system, and the maximum PV system size could be set to some smaller number.) To see a significant improvement in solve time, users could fix the system size by setting the minimum and maximum system size to the same value and assess the economics of that system, though this option means that REopt is not optimizing the system size.
• Reduce number of technologies being evaluated concurrently: This helps reduce the number of decision variables and thus solve times.
• Reduce use of capabilities requiring binary variables: Binary variables in the optimization tend to slow down solve times, regardless of solver. Examples of capabilities formulated with binary variables in REopt include net metering limits, resilience analyses, tiered utility rates, and minimum turndown on generators and CHP technologies. Avoiding or zeroing out these inputs can help speed up solve times.
• Reduce use of capabilities leveraging challenging constraint formulations (indicator constraints, big M constraints): Some capabilities in REopt, namely net metering, resilience, and tiered energy rates, use formulations of constraints known as “indicator constraints” or “big M” constraints (see below under “Solver selection” for additional information). Evaluations involving these capabilities may be slow to solve.

Adjust solver settings: Users can adjust the optimization solver selection, maximum solve time, and optimality tolerance to improve performance. In the web tool, these three optimization settings can be modified in the “Solver Settings” section at the bottom of the “Site and Utility” inputs accordion. In the API, these three inputs can be specified by Settings.solver_name (note: this input is not case sensitive), Settings.timeout_seconds, and Settings.optimality_tolerance, respectively.

• Optimality tolerance: To help speed up solve times, users can increase the optimality tolerance for the optimized scenario, albeit potentially risking the solver stopping at a solution before a more-optimal solution was found. Once a solution has been identified using a larger tolerance, the user may bound the minimum and maximum technology sizes around the REopt-identified sizes and then re-run with a reduced tolerance.
• Maximum solve time: To accommodate slower solve times, we have increased REopt’s default and maximum solve times to 600 seconds and 1,200 seconds, respectively. Users can increase the solve time, up to 1,200 seconds, to reduce timeout errors. Note that a user can close out of a results URL while the optimization is still solving and navigate back to that URL to obtain results once complete. Additionally, logging in to a REopt user account allows users to access all evaluations completed while logged in in a “Saved Evaluations” page.
• Solver selection: REopt now uses HiGHS as its default solver, because it tends to have the fastest solve times based on our benchmarking. However, for some problems Cbc or SCIP may solve faster, for instance Cbc was observed to solve faster for some “financial” and “clean energy” scenarios in our initial testing. Users can try running the same analysis with different solvers to explore solve time improvements with different solvers. Note: This updated version of REopt API and web tool uses re-written indicator constraints as big M constraints for compatibility with HiGHS, Cbc, and SCIP.