BOOST-style microgrid sizing reimplementation

This is a clean reimplementation of the core idea in BOOST: Microgrid Sizing using Ordinal Optimization:

1. evaluate many candidate (battery size, PV size) designs with a simple LP;
2. rank them by annual operating cost + annualized investment cost;
3. re-evaluate only the top-s designs with a more accurate MILP that adds the diesel minimum-output logic.

What is included

• boost_sizing/dispatch.py
  Weekly dispatch solver:

  • LP = simple model
  • MILP = accurate model with diesel on/off binary

• boost_sizing/yearly.py
  Rolls the weekly dispatch solver across the full year with SOC carryover from week to week.

• boost_sizing/oo.py
  Implements:

  • the N formula from the paper
  • the alignment-probability formula
  • automatic choice of s

• boost_sizing/boost.py
  End-to-end experiment runner:

  • sample N designs
  • solve LP for each
  • rank designs
  • solve MILP on top s
  • export CSVs, PNG plots, and JSON summary

• boost_sizing/baselines.py
  Two comparison baselines:

  • greedy dispatch
  • approximate DP dispatch

• boost_sizing/synthetic_data.py
  Creates an hourly synthetic year of:

  • load
  • solar capacity factor
  • time-of-use grid price

Why this version is slightly more explicit than the paper

The paper is short, so a few engineering details have to be made explicit in code. This implementation therefore adds or fixes the following:

• battery charge/discharge power limits
  The paper shows SOC dynamics and SOC bounds, but does not visibly spell out explicit battery power-rate constraints in the short formulation. Here they are added directly.

• weekly SOC carryover
  The paper says a new optimization is solved every week. This code carries the final SOC from one week into the next.

• annualized costs
  PV and battery investment costs are annualized using the standard capital recovery factor.

• synthetic demo data by default
  This keeps the code runnable out of the box. You can swap in real hourly CSVs easily by replacing the generate_synthetic_year(...) call with your own loader.

Default modeling choices

These are editable in boost_sizing/config.py.

• battery max charge power = 0.10 * battery_kwh kW
• battery max discharge power = 0.10 * battery_kwh kW
• SOC bounds = 10% to 90%
• charge / discharge efficiencies = 95%
• diesel minimum output = 150 kW
• diesel maximum output = 1500 kW
• grid maximum import = 2500 kW

The 10%-per-hour battery power limit matches a conservative engineering default and can be changed in one place.

Quick start

python demo_quickstart.py

This runs an 8-week smoke test and writes outputs under:

results/quickstart/

Full(er) run

python run_experiment.py --out_dir results/run_full

For the default 10x10 design grid, the theoretical ordinal-optimization formula gives about 90 candidate designs for the paper's P=99%, alpha=5% setting.

Main outputs

• lp_rankings.csv
  LP-screened designs

• accurate_top_s.csv
  top-s designs after accurate MILP re-evaluation

• best_schedule.csv
  yearly schedule for the best accurate design

• baseline_comparison.csv
  LCOE comparison of BOOST vs. DP vs. greedy on the selected best design

• summary.json
  overall experiment summary

• figs/*.png
  rank-stability and LCOE plots

Notes on faithfulness

This is method-faithful, not a byte-for-byte clone of the original public codebase. In particular:

• it uses SciPy / HiGHS instead of the original repository's implementation stack;
• it uses synthetic demo data unless you provide your own real data;
• the DP baseline is intentionally approximate and documented as such.

That said, the core paper logic is preserved:

• simple LP screening,
• top-s refinement with a more accurate MILP,
• ordinal-optimization formulas for N and s,
• total-cost ranking,
• LCOE reporting.

Suggested next upgrades

1. swap synthetic data for real hourly load / irradiance / tariff CSVs;
2. replace the approximate DP baseline with the exact DP from the author's earlier microgrid work;
3. add battery degradation and curtailment penalties;
4. add explicit grid outage/intermittency modeling if desired.