RAChronoOps

RAChronoOps provides a systematic storage-dispatch fidelity ladder for sequential Monte Carlo resource adequacy (RA), building on prior work showing that storage dispatch affects adequacy metrics (Gonzato et al. 2023; PRAS/Stephen 2021). The project compares naive proactive heuristics, PRAS/Evans-style emergency-only dispatch, event-window LP, full-year ED, PCM-ED, and PCM-UCED under common random numbers on the public RTS-GMLC single-zone test system. It quantifies which storage dispatch approximations preserve EUE, LOLH, CVaR-EUE, and runtime performance, and introduces a storage-energy sufficiency-bound diagnostic that explains when simpler methods recover full-year ED EUE.

Method hierarchy

Label	Paper name	Method	Optimization	Status
MC-NoStorage	Traditional MC	Traditional hourly sequential MC without storage	None	Implemented — no-storage baseline
RA-1a / M1	Naive Storage MC	Sequential MC + naive peak-shaving storage heuristic	None	Implemented — cautionary baseline
RA-1b / M1b	Reserve-Floor MC	Sequential MC + reserve-aware storage heuristic	None	Implemented and validated
RA-1c / M1c	Emergency-Only MC	Sequential MC + PRAS/Evans-style emergency-only dispatch (system-surplus charging)	None	Implemented and validated
M1c_VREOnly	VRE-Surplus MC	Same as M1c but charges from VRE surplus only	None	Implemented — appendix sensitivity
RA-1d / M1d	Risk-Hour MC	Sequential MC + risk-hour allocation heuristic (earliest_first / largest_first)	None	Implemented — within-event allocation study
RA-2 / M2	Event-Window LP-MC	Sequential MC + screened event-window LP near risk periods (proposed scalable method)	Small LPs	Implemented and validated
RA-3 / M3	Full-Year ED-MC	Sequential MC + full-year ED LP per scenario	LP (Gurobi)	Implemented — benchmark
HOPE-ED	PCM-ED	Full-year HOPE ED LP (PCM economic dispatch mode)	LP (Gurobi)	Validated — matches M3
HOPE-UC / M4	PCM-UCED	Full-year HOPE UC/PCM with real Pmin/ramp/startup/min-up/down	MILP (Gurobi)	Validated — N=5 and N=20

All implemented methods use common random numbers: one shared ScenarioSet of thermal outage draws is passed to every method, so reliability metric differences are attributable to dispatch strategy alone.

Quick start

# 1. Install Julia dependencies
julia --project=. -e "using Pkg; Pkg.instantiate()"

# 2. Download RTS-GMLC data (requires git on PATH)
julia --project=. scripts/00_get_rts_gmlc_data.jl

# 3. Build the processed single-zone dataset
julia --project=. scripts/01_build_single_zone_rts.jl

# 4. Build VRE experiment cases
julia --project=. scripts/06_build_experiment_cases.jl

# 5. Build no-storage case variants
julia --project=. scripts/33_build_no_storage_cases.jl

# 6. No-storage MC vs ED baseline comparison (N=20)
julia --project=. scripts/34_compare_no_storage_classic_vs_ed.jl \
  --cases VRE120_base,VRE120_wind_hvy --n-scenarios 20 --seed 42

# 7. Export HOPE full-year cases (5 scenarios, ED + UC)
julia --project=. scripts/25_build_hope_full_year_cases.jl \
  --case VRE120_base --modes ED,UC --n-scenarios 20 --seed 42

# 8. HOPE no-storage four-model comparison (requires script 29 + 27 first)
julia --project=. scripts/36_compare_nostorage_hope_uc_n5.jl

# 9. Wind-heavy five-model storage-enabled HOPE-UC comparison
julia --project=. scripts/37_compare_wind_hvy_hope_uc_n5.jl

Current completed experiments

The full experiment sequence is complete. Clean narrative: traditional sequential MC is valid without storage; storage introduces intertemporal SOC coupling that makes reliability estimates sensitive to dispatch assumptions; M1c and M2 recover M3/HOPE EUE at much lower runtime; HOPE-UC mainly changes LOLH/event timing, not EUE. A storage-energy sufficiency bound explains why EUE convergence across M1c/M1d/M2/M3 is observed in the tested RTS-GMLC cases: in these configurations, the binding constraint is storage energy availability (MWh) rather than dispatch complexity.

1. No-storage MC validation

MC-NoStorage = M3-NoStorage exactly for VRE120_base and VRE120_wind_hvy at N=20. For VRE120_base_nostorage at N=5, all four models agree: MC-NoStorage = M3-NoStorage = PCM-ED-NS (HOPE-ED-NoStorage) = PCM-UCED-NS (HOPE-UC-NoStorage) (ΔEUE = 0.00 MWh, ΔLOLH = 0.0 h).

Interpretation: traditional sequential MC is valid in the classic no-storage RA setting. Without storage there is no intertemporal state variable, so LP and MILP dispatch collapse to the same feasible set as the classical capacity check.

Scripts: scripts/33_build_no_storage_cases.jl, scripts/34_compare_no_storage_classic_vs_ed.jl, scripts/36_compare_nostorage_hope_uc_n5.jl

Results: results/no_storage_comparison/, results/nostorage_hope_uc_comparison/

2. Storage-aware MC method comparison

Model	LOLH vs M3	EUE vs M3	Runtime
M1 (naive heuristic)	High bias	High bias	~1 s/scenario
M1b (reserve-aware)	Reduced bias	Reduced bias	~1 s/scenario
M1c (emergency-only)	Matches M3	Matches M3	~1–2 s/scenario
M2 (event-window LP)	Within 0.2 h	Matches M3	5–10 s/scenario
M3 (full-year ED)	—	— (benchmark)	~9–10 s/scenario

M1 / M1b overestimate reliability risk (naive peak-shaving depletes storage before shortage events). M1c and M2 recover M3 EUE/CVaR closely.

Recommended M2 config: risk_margin_mw=1000, window_buffer_hours=48.

Scripts: scripts/14_run_ra1b_validation.jl, scripts/16_run_vre_method_comparison.jl, scripts/30_compare_all_models_hope_n5.jl

3. PCM full-year validation (PCM-ED and PCM-UCED)

PCM-ED (HOPE-ED internally) matches M3 once ED-mode ramp constraints are disabled (ΔEUE < 1 MWh). PCM-UCED (HOPE-UC internally) uses real Pmin, ramp rates, startup costs, and min up/down times from the RTS-GMLC dataset.

In tested cases (VRE120_base N=20; VRE120_wind_hvy N=5), PCM-UCED mainly changes LOLH/event timing, not EUE:

Case	Model	LOLH (h)	EUE (MWh)	Runtime (s)
VRE120_base	PCM-ED	6.2	2,479	2,356
VRE120_base	PCM-UCED	7.2	2,479	11,438
VRE120_wind_hvy	PCM-ED	3.8	1,113	588
VRE120_wind_hvy	PCM-UCED	4.2	1,113	2,712

Runtime note: PCM-UCED is 4–5× slower than PCM-ED with zero EUE benefit in the tested cases. UC is worth running only when storage is present (storage SOC is the intertemporal link that activates UC timing effects).

Scripts: scripts/25_build_hope_full_year_cases.jl, scripts/29_run_hope_n5_pilot.jl, scripts/27_collect_hope_results.jl, scripts/37_compare_wind_hvy_hope_uc_n5.jl

Results: results/wind_hvy_hope_uc_comparison/, results/hope_wind_hvy_n5_pilot/

4. Storage-energy sufficiency bound (theoretical diagnostic)

The sufficiency bound derives a per-scenario ceiling on EUE reduction from storage, given surplus energy available in a 72-hour lookback window before each shortage event:

coverage_bound = min(pre_event_EUE,
                     feasible_discharge_energy,   -- SOC × η_dis after charging
                     power_limited_coverage)       -- Σ min(shortfall[h], power_mw)

Case	Pre-storage EUE	Bound/M3 EUE	Sufficiency ratio
VRE120_base	31,017 MWh	2,479 MWh	0.941
VRE120_wind_hvy	15,801 MWh	648 MWh	0.972

The bound matches M3 EUE exactly per scenario (ΔEUE = 0.00 MWh). Binding constraint in tested cases: storage energy (MWh), not power (MW).

Key insight: M1c, M1d_earliest, M2, and M3 all achieve the bound because they satisfy its two sufficient conditions (charge from surplus; discharge only at shortfall hours). M1/M1b violate condition 2 by proactively discharging, depleting SOC before shortage events and moving away from the bound.

Why LOLH can differ even when EUE matches: EUE is set by the storage energy budget; LOLH counts hours with any positive load shed. M1d_largest reallocates discharge to the highest-shortfall hours first, leaving smaller shortfall hours partially served (more LOLH, same EUE). LP degeneracy and HOPE-UC commitment constraints produce the same effect for different reasons.

Scripts: scripts/38_compare_m1d_storage_heuristics.jl, scripts/39_storage_energy_sufficiency_bound.jl

Results: results/m1d_storage_heuristic_comparison/, results/storage_energy_sufficiency_bound/

Documentation: docs/storage_energy_sufficiency_bound.md

Full numeric results and per-script output paths: docs/results_index.md and docs/current_findings_synthesis.md.

Archived diagnostics

The following early experiments informed the current design but are no longer the primary results:

Rolling-window M2 (M2RollingWindow.jl): solves one LP per hour per scenario — no runtime advantage over M3. Replaced by the event-window M2 (M2EventWindowLP.jl).
HOPE stress-week validation (Phase E placeholder): the project moved to full-year HOPE ED/UC, making the stress-week-only approach unnecessary.
RA-1a / M1 diagnosis: priority-2 proactive discharge depletes SOC to zero before 100% of shortage events; motivates M1b and M1c.

See docs/experiment_archive.md for full diagnostic results.

Preparing the full RTS-GMLC dataset

The official RTS-GMLC dataset is hosted at https://github.com/GridMod/RTS-GMLC.

# Clone RTS-GMLC into data_raw/RTS-GMLC/ (idempotent)
julia --project=. scripts/00_get_rts_gmlc_data.jl

# Build the 8760-hour single-zone processed dataset
julia --project=. scripts/01_build_single_zone_rts.jl

# Optionally verify: print system summary + write results/data_summary/ CSVs
julia --project=. scripts/04_summarize_processed_data.jl

What the builder does

BuildRTSSingleZone.jl aggregates the RTS-GMLC three-area system into a single copper-plate zone:

Step	Detail
Generator fleet	133 units: 73 thermal + 60 VRE (4 WIND, 25 PV, 31 RTPV)
Excluded types	HYDRO (19), ROR (1), CSP (1), STORAGE (1), SYNC_COND (3) — logged as warnings
Heat rate	`HR_avg_0` column (BTU/kWh) ÷ 1000 → MMBTU/MWh; `variable_cost = VOM + HR × fuel_price`
Load	`DAY_AHEAD_regional_Load.csv` columns 5+ (three regions) summed per hour, truncated to 8760 h
Wind CF	`WIND/DAY_AHEAD_wind.csv` MW ÷ total wind capacity (2507.9 MW), clamped to [0, 1]
Solar CF	`PV/DAY_AHEAD_pv.csv` + `RTPV/DAY_AHEAD_rtpv.csv` MW ÷ total solar capacity (2715.9 MW)
Storage	One aggregate 4-hour battery: 10% of peak load power, η = √0.90 per half-trip
Validation	Strict (8760 h required) when RTS data detected; lenient for synthetic fallback

Expected output (RTS-GMLC)

After running script 01, results/data_summary/system_summary.csv should show:

Metric	Value
n_hours	8760
Peak load	8191.8 MW
Annual load	37,561 GWh
Thermal capacity	8076 MW (73 units)
Wind capacity	2508 MW (CF 32.4%)
Solar capacity	2716 MW (CF 24.7%)
Storage	819 MW / 3276 MWh (4 h)

Synthetic fallback

If RTS-GMLC data is absent (e.g., CI or first checkout without step 00), the builder writes a deterministic 168-hour system so that tests and scripts still run.

Building experiment cases

scripts/06_build_experiment_cases.jl writes one subfolder per case under data_processed/cases/.

Experiment group	Scale factors applied	Status
Load scaling (×5)	`load_scale` ∈ {1.00, 1.05, 1.10, 1.15, 1.20}	Diagnostic — completed
Storage matrix (×12)	`storage_power_pct_peak` ∈ {5%, 10%, 20%} × `storage_duration_hours` ∈ {2, 4, 8, 12}	Diagnostic — completed
VRE penetration/profile (×6)	wind/solar scale pairs — see docs/redesigned_experiment_plan.md	Main experiment — completed

julia --project=. scripts/06_build_experiment_cases.jl

Reliability metric taxonomy

All implemented methods compute a common set of metrics stored in MetricsResult (see src/metrics/ReliabilityMetrics.jl).

Frequency of shortfall

Metric	Field	Definition
LOLH	`lolh`	Mean loss-of-load hours per year (h/yr)
LOLP	`lolp`	Loss-of-load probability = LOLH / 8760 (direct normalization in hourly sequential simulation)
LOLE days	`lole_days`	Mean days per year with ≥ 1 shortage hour

Energy not served

Metric	Field	Definition
EUE	`eue`	Expected unserved energy (MWh/yr)
nEUE	`neue`	EUE / annual load energy (fraction; report in ppm = × 10⁶)

Shortage event structure

Metric	Field	Definition
Event count	`n_shortage_events`	Mean number of distinct contiguous shortage events per scenario
Mean duration	`mean_shortage_duration`	Mean event duration (h)
Max duration	`max_shortage_duration`	Maximum event duration across all scenarios (h)
p95 duration	`p95_shortage_duration`	95th percentile of event duration pooled across scenarios (h)

Shortfall severity

Metric	Field	Definition
Max shortfall	`max_shortfall`	Maximum single-hour load shed across all scenario-hours (MW)
Mean shortfall	`mean_shortfall_when_shedding`	Mean load shed conditional on shed > 0 (MW)

Tail risk (scenario distribution)

Metric	Field	Definition
p50/p90/p95/p99 EUE	`p{q}_scenario_eue`	Percentiles of the per-scenario EUE distribution (MWh)
CVaR-EUE	`cvar_eue`	Mean of the top 5% of per-scenario EUEs (MWh)

Monte Carlo uncertainty

Metric	Field	Definition
LOLH CI95	`lolh_ci95_halfwidth`	1.96 × std(per-scenario LOLH) / √N (h)
EUE CI95	`eue_ci95_halfwidth`	1.96 × std(per-scenario EUE) / √N (MWh)

Scope

The study evaluates centralized adequacy-oriented storage dispatch. The following are explicitly outside scope:

Merchant storage behavior: no bidding, capacity withholding, strategic behavior, or decision-dependent storage availability.
Operating reserves and ancillary services: no reserve requirements or forecast-error-driven unit commitment beyond traditional RA assumptions.
Network constraints: single copper-plate zone; multi-area systems are future work.
Transmission network, demand response, hydro water budget.
Imperfect foresight and investment re-optimization.

These exclusions align with traditional RA conventions and define the boundary of applicability. Unit commitment is introduced only in PCM-UCED (HOPE-UC) as the high-fidelity validation benchmark.

Repository layout

RAChronoOps/
  Project.toml
  docs/
    current_findings_synthesis.md  # key findings across all experiments
    redesigned_experiment_plan.md  # current experiment design
    results_index.md               # index of result folders
    storage_energy_sufficiency_bound.md  # theory note on the sufficiency bound
    experiment_archive.md          # completed diagnostic results
    hope_full_year_case_preparation.md  # HOPE case export details
    next_experiment_design_nostorage_and_real_uc.md
    ra2_n20_validation_memo.md
    vre_method_comparison_memo.md
    m1c_charging_assumption_memo.md
    ra1b_validation_memo.md
  src/
    RAChronoOps.jl
    utils/
      Config.jl
      IO.jl
    data/
      GetRTSGMLCData.jl
      LoadData.jl
      BuildRTSSingleZone.jl
      ExportHOPECase.jl
    scenarios/
      SequentialOutages.jl
    models/
      McNoStorage.jl               # classical no-storage MC baseline
      M1RuleBasedStorage.jl        # RA-1a: naive peak-shaving heuristic
      M1bReserveAwareStorage.jl    # RA-1b: reserve-aware heuristic
      M1cEmergencyOnlyStorage.jl   # RA-1c: emergency-only heuristic
      M1cVREOnlyCharge.jl          # M1c_VREOnly: VRE-surplus charging variant
      M1dRiskHourAllocation.jl     # RA-1d: risk-hour allocation heuristic
      M2EventWindowLP.jl           # RA-2: event-window LP
      M2RollingWindow.jl           # rolling-window LP (archived diagnostic)
      M3EDDispatch.jl              # RA-3: full-year ED LP benchmark
    metrics/
      ReliabilityMetrics.jl
    experiments/
      RunExperiment.jl
  scripts/
    # ── data preparation ──────────────────────────────────────────────────
    00_get_rts_gmlc_data.jl
    01_build_single_zone_rts.jl
    04_summarize_processed_data.jl
    05_smoke_test_full_rts_data.jl
    06_build_experiment_cases.jl
    # ── main comparison scripts ───────────────────────────────────────────
    30_compare_all_models_hope_n5.jl     # M1c/M2/M3/HOPE-ED on base case N=5
    33_build_no_storage_cases.jl         # build <case>_nostorage variants
    34_compare_no_storage_classic_vs_ed.jl  # MC-NoStorage vs M3-NoStorage
    36_compare_nostorage_hope_uc_n5.jl   # four-model no-storage HOPE-UC check
    37_compare_wind_hvy_hope_uc_n5.jl    # five-model wind-heavy HOPE-UC check
    38_compare_m1d_storage_heuristics.jl # M1c/M1d/M2/M3 within-event allocation study
    39_storage_energy_sufficiency_bound.jl # theoretical sufficiency bound diagnostic
    # ── HOPE export and run ───────────────────────────────────────────────
    25_build_hope_full_year_cases.jl     # export HOPE full-year case folders
    27_collect_hope_results.jl           # collect HOPE output metrics
    29_run_hope_n5_pilot.jl              # run HOPE cases via Julia subprocess
    # ── diagnostic scripts (archived) ────────────────────────────────────
    09_calibrate_load_scaling.jl
    10_run_storage_matrix.jl
    11_debug_storage_cases.jl
    12_run_selected_storage_validation.jl
    13_debug_m1_storage_sensitivity.jl
    14_run_ra1b_validation.jl
    16_run_vre_method_comparison.jl
  test/
    runtests.jl
    test_storage.jl
    test_m1d.jl
    test_outages.jl
    test_power_balance.jl
    test_common_scenarios.jl
  results/
    # main experiment results
    no_storage_comparison/         # MC-NoStorage vs M3-NoStorage (N=20)
    nostorage_hope_uc_comparison/  # four-model no-storage HOPE-UC check
    wind_hvy_hope_uc_comparison/   # five-model wind-heavy HOPE-UC check
    hope_wind_hvy_n5_pilot/        # HOPE run status + metrics for wind-hvy
    hope_nostorage_n5_pilot/       # HOPE run status + metrics for nostorage
    full_model_comparison_with_hope/  # M1c/M2/M3/HOPE-ED on base case
    vre_method_comparison/         # M1/M1b/M1c/M2/M3 across VRE cases
    m1d_storage_heuristic_comparison/  # M1c/M1d/M2/M3 within-event allocation
    storage_energy_sufficiency_bound/  # theoretical sufficiency bound
    # diagnostic results (archived)
    storage_matrix/
    storage_validation/
    m1_debug/
    load_scaling/

Name		Name	Last commit message	Last commit date
config		config
configs		configs
data_processed		data_processed
data_raw		data_raw
docs		docs
exports/hope_model_cases		exports/hope_model_cases
figures		figures
logs		logs
paper @ 7dc57ef		paper @ 7dc57ef
results		results
scripts		scripts
src		src
test		test
.gitignore		.gitignore
.gitmodules		.gitmodules
Project.toml		Project.toml
README.md		README.md
calib09_err.txt		calib09_err.txt
calib09_out.txt		calib09_out.txt
calib_err.txt		calib_err.txt
calib_out.txt		calib_out.txt
julia_m3_err.txt		julia_m3_err.txt
julia_m3_out.txt		julia_m3_out.txt
test_run_output.txt		test_run_output.txt

Navigation Menu

Repository files navigation

RAChronoOps

Method hierarchy

Quick start

Current completed experiments

1. No-storage MC validation

2. Storage-aware MC method comparison

3. PCM full-year validation (PCM-ED and PCM-UCED)

4. Storage-energy sufficiency bound (theoretical diagnostic)

Archived diagnostics

Preparing the full RTS-GMLC dataset

What the builder does

Expected output (RTS-GMLC)

Synthetic fallback

Building experiment cases

Reliability metric taxonomy

Frequency of shortfall

Energy not served

Shortage event structure

Shortfall severity

Tail risk (scenario distribution)

Monte Carlo uncertainty

Scope

Repository layout

Further reading

About

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

RAChronoOps

Method hierarchy

Quick start

Current completed experiments

1. No-storage MC validation

2. Storage-aware MC method comparison

3. PCM full-year validation (PCM-ED and PCM-UCED)

4. Storage-energy sufficiency bound (theoretical diagnostic)

Archived diagnostics

Preparing the full RTS-GMLC dataset

What the builder does

Expected output (RTS-GMLC)

Synthetic fallback

Building experiment cases

Reliability metric taxonomy

Frequency of shortfall

Energy not served

Shortage event structure

Shortfall severity

Tail risk (scenario distribution)

Monte Carlo uncertainty

Scope

Repository layout

Further reading

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages