In build · CEMeridian

By Compounding Energy

In build Platform infrastructure — replaces our HiGHS dependency across CEGridSight, CEForesight, CENovaSage, CE BESS Arbitrage, and CEDeris.

A native Rust LP, MIP, and QP solver — built for the problems the rest of the platform actually solves.

CEGridSight runs on HiGHS today. It works, it's fast enough, and it's the right starting point — but it isn't ours, and the structure of a 27-zone, 480-period unit-commitment MILP isn't the structure a generic backend exploits well. CEMeridian is the in-house replacement: native Rust, dispatch-aware decomposition, warm-started from production shadows, and instrumented to debug the dispatch — not just the solver.

Get on the early-access list CEAtlas Expert pricing

Why this exists

A solver should know what problem it's solving.

Generic LP/MIP backends — HiGHS, CBC, the commercial ones — treat every problem as an arbitrary sparse matrix. Power-system dispatch isn't arbitrary: it has a per-zone block structure, predictable temporal coupling, and a small set of binary decisions that explain almost all of the integrality gap. CEMeridian is built around that structure rather than against it, with the platform's own warm-start data as a first-class input rather than an afterthought.

What it does

Three pillars.

LP, MIP, and QP in one stack

Linear, mixed-integer, and quadratic programming under a single Rust core. The QP path matters specifically for risk-aware bidding (CE BESS Arbitrage) and for AC-OPF screening (CEForesight); doing it in one codebase avoids three unrelated solver dependencies.

Dispatch-aware decomposition

Benders and Lagrangian splits aligned to the natural per-zone and per-time-block structure of the dispatch MILP — exploiting the structure that's actually there, rather than the structure a generic preprocessor can guess.

Warm-started by construction

Each new run warm-starts from the previous 6-hour MILP solution — basis, integer assignments, and binding-cut history all carried forward. The expected delta between consecutive runs is small; we exploit that explicitly rather than re-solving from scratch.

Methodology

For the quants in the room.

Plain-English first; this section for anyone vetting the maths. Skip if you're not building the optimiser yourself.

LP core

Dual simplex + interior point.

Native Rust dual simplex with bound-flipping ratio test and steepest-edge pricing for warm-started re-optimisation; interior-point (Mehrotra predictor-corrector) for cold solves on large LPs. Crossover to a basic solution provided when downstream MIP needs it.

MIP

Branch-and-cut with structural cuts.

Branch-and-cut on top of the LP core, with cut families derived from the unit-commitment structure: tight Carrion–Arroyo formulation, Damci-Kurt cuts, and ramp-and-min-up/down cuts generated structurally rather than learned. Strong branching restricted to commitment binaries that empirically dominate the gap.

Active-set with rank-one updates.

Active-set QP for warm-started CVaR-aware battery bidding; convex Hessians from the risk objective, sparse linear constraints from the energy/SoC structure. Rank-one Cholesky updates as the active set evolves; degeneracy handled by Bland's rule fallback.

Decomposition

Benders + temporal block-coordinate.

Benders cuts on inter-zonal coupling constraints; block-coordinate descent on temporal blocks of representative weeks for the capacity-expansion problem. Master and subproblem solves run in parallel where the cut graph allows; cuts cached per problem hash and reused across consecutive runs.

Instrumentation

Per-block, per-cut, per-constraint traces.

Every solve emits a structured trace — solve time per block, cuts added per round, dual values for every binding constraint, basis turnover. We can debug why a dispatch run priced badly, not just why it took 8 minutes instead of 3.

Interface

Python API compatible with HiGHS.

Drop-in replacement for the HiGHS Python interface used across the platform today. Switch the backend per environment variable; all CEGridSight, CEForesight, CENovaSage, CE BESS Arbitrage, and CEDeris workloads run unchanged on either solver during the migration.