Brazil Zonal Curtailment Simulation

This repo builds demand/generation forecasts, runs a zonal DC-OPF with ENS/dump handling, and summarizes curtailment across multiple scenarios.

Data prep (run once)

1. CSV preprocessing (csv_preprocessing/): clean/align ONS generation, weather, and demand inputs. Key outputs:
   • data/merged_generation_weather_v2.csv (plant gen + weather, daily)
   • data/demand_data/demand_projection_clean.csv (state/subsystem monthly MWh)
2. Network build (already committed): relaxed 5-bus net at models/pandapower_snapshots/brazil_network_zonal_5bus_relaxed.json.

Core workflow

# Demand forecast (scenarios: inferior/referencia/superior)
python forecast_demand_from_table.py --scenario referencia --output data/demand_data/demand_projection_2025_2028.csv

# Generation forecast (auto-eval both models; uses HGB by default)
python forecast_generation_ml2.py --val-leakage --output results/generation_forecast_2025_2028.csv

# Curtailment sim (monthly, zonal DC-OPF)
python run_forecast_curtailment_sim.py \
  --net-json models/pandapower_snapshots/brazil_network_zonal_5bus_relaxed.json \
  --demand data/demand_data/demand_projection_2025_2028.csv \
  --gen results/generation_forecast_2025_2028.csv \
  --n-trials 200 \
  --line-loading-percent 100 \
  --out results/curtailment_simulations.csv

# Summaries + plots
python summarize_curtailment_sim.py --input results/curtailment_simulations.csv

Optional helpers

• Batch scenarios: bash run_scenario_batch.sh (runs inferior/referencia/superior end-to-end).
• Sensitivity sweep: python run_sensitivity.py (line loading × slack × scenarios) then python summarize_sensitivity.py.
• Quarterly plot: after a historic baseline + future runs, python plot_quarterly_curtailment.py --historic <historic_monthly_detailed.csv>.

Key assumptions

• Demand forecasts are average MW per month derived from growth targets + seasonal factors (2020–2024 history).
• Generation forecasts are monthly avg MW per subsystem via AR+seasonal models (ridge/HGB) with weather features and optional noise; --val-leakage uses actual lags for validation.
• OPF is monthly DC, zonal 5-bus; ENS/dump modeled as high-cost injections; slack optional. No time coupling/storage.
• Historical gen with daily timestamps is treated as MWh/day and divided by 24 to get MW.

Repro tips

• Use python3 in your venv with required deps (pandas, numpy, sklearn, pandapower, matplotlib).
• Seeds are exposed (--seed); set forecast noise stds to 0 for deterministic runs.
• Results are gitignored (results/).