Deep Wind Model Dispatch

This repository contains Summer 2026 REU work on forecast-aware dispatch for a hybrid wind farm with compressed-air-energy-storage-like constraints. The clean paper-facing reproduction package is organized here:

Summer 2026 REU/

The project is now organized around one primary benchmark:

100-MW Constant-Output Baseload Benchmark

Wind-only/no-storage output is still reported, but only as secondary reference information. The 100 MW benchmark uses the same 100 MW / 10 h / 1,000 MWh CAES configuration as the dispatch experiments.

Reproduction Ladder

Run one command in each folder:

Step	Folder	What it proves
0	Summer 2026 REU/100 MW baseload	Builds the 100 MW benchmark and QA checks
1	Summer 2026 REU/causal ridge regression	Chooses the best power forecast by RMSE
2	Summer 2026 REU/rolling horizon	Tests deterministic Gurobi planning horizons
3	Summer 2026 REU/different scenarios	Tests 1/3/5/7/10 forecast scenarios
4	Summer 2026 REU/oracle upper bound	Reports daily and hourly perfect-information oracle references

Every folder has:

EXPERIMENT_KNOBS.py
RUN_*.py
code/
results/
figures/

Change the knobs file, run the RUN_*.py command, then read results/current_run_from_knobs/.

Current Paper-Facing Results

Result	Current value
Best forecast model	causal lag/ridge forecast
Best forecast RMSE	21.24 MW
Best deterministic rolling-horizon case	48 h
Deterministic COVE reduction vs 100 MW benchmark	20.63%
Best scenario case	3 scenarios
Scenario COVE reduction vs 100 MW benchmark	40.18%
Scenario revenue gain vs 100 MW benchmark	67.16%
Daily-replan oracle reference	168 h, 40.87% COVE reduction
Hourly-replan oracle reference	168 h, 40.85% COVE reduction

Step 2 48 h vs Step 3 1 Forecast

These are intentionally different rows, so their percentages should not match.

Value	Source	Meaning
20.63%	Summer 2026 REU/rolling horizon/results/current_run_from_knobs/forecast_dispatch_summary.csv	Step 2 deterministic horizon sweep: one causal forecast path, 48 h planning horizon, 24 h execution/replanning, and the 75 MW direct-reserve policy
37.23%	Summer 2026 REU/different scenarios/results/current_run_from_knobs/uncertainty_aware_summary.csv	Step 3 one-forecast reference inside the scenario runner: current-hour nowcast plus causal ridge future, nowcast-gated recourse, and first-action execution

Use Step 2 to compare planning horizons. Use Step 3 to compare one forecast against 3, 5, 7, and 10 scenarios.

Common Storage Setup

Setting	Value
Storage power	100 MW
Duration	10 h
Capacity	1,000 MWh
SoC bounds	200-1,000 MWh
Initial SoC	600 MWh
RTE	55%, discharge-side
Grid export cap	249 MW
Grid charging	Not allowed

Main Commands

cd "/Users/davidvalenta/deep-wind-model-dispatch/Summer 2026 REU/100 MW baseload"
../../venv/bin/python RUN_0_100MW_BASELOAD.py

cd "/Users/davidvalenta/deep-wind-model-dispatch/Summer 2026 REU/causal ridge regression"
../../venv/bin/python RUN_1_FORECAST_RMSE.py

cd "/Users/davidvalenta/deep-wind-model-dispatch/Summer 2026 REU/rolling horizon"
../../venv/bin/python RUN_2_ROLLING_HORIZON.py

cd "/Users/davidvalenta/deep-wind-model-dispatch/Summer 2026 REU/different scenarios"
../../venv/bin/python RUN_3_SCENARIO_COMPARISON.py

cd "/Users/davidvalenta/deep-wind-model-dispatch/Summer 2026 REU/oracle upper bound"
../../venv/bin/python RUN_4_ORACLE_UPPER_BOUND.py

Run the audit:

cd "/Users/davidvalenta/deep-wind-model-dispatch/Summer 2026 REU"
../venv/bin/python AUDIT_DATA_CONFIG.py

Current audit result: 16/16 active hourly CSV files pass the common 100 MW / 10 h CAES checks.