mkaracelebi/SequentialContingencyFinder_DB

Experience Database — README

Overview

rl_experience.db and rl_experience_test.db are SQLite databases that serve as persistent caches for the RL training pipeline (rl_experience_store.py). Each row stores the result of one RL transition — a unique (scenario_id, previous_lines, action_line_index) triple — including its dynamic-impact severity (R_D, computed via Dynawo), relaxed-OPF severity (R_O, computed via IPOPT), and the resulting state feature vector. Computing one such triple takes ~2–3 s for R_D and ~0.3–0.5 s for R_O; the database means each triple is computed exactly once and then served from disk forever, across all processes and training runs.

alpha is deliberately not stored. Because R_total = alpha * R_D + (1 - alpha) * R_O, any blend can be recomputed at query time from the cached R_D/R_O pair with no re-simulation.

---

Files

File	Role	Scenario IDs
rl_experience.db	Training cache	0 – 299 (random dispatch seeds)
rl_experience_test.db	Held-out evaluation cache	10 000 – 10 099 (never seen during training)

---

Schema

Both databases share the same two-table schema.

transitions (primary table)

Each row is one cached (scenario, prefix, action) triple.

Column	Type	Description
scenario_id	INTEGER	Dispatch seed passed to DispatchGenerator(seed=scenario_id)
previous_lines	TEXT	Comma-separated line indices already tripped before this action (empty string "" for depth-0 / first fault)
action_line_index	INTEGER	The pandapower line index tripped at this step
step	INTEGER	Cascade depth of this action (0 = first fault, 1 = second, 2 = third)
R_D	REAL	Dynamic-impact severity from Dynawo transient simulation. Set to 5000.0 if Dynawo or the pre-fault power flow fails to converge
R_D_max	REAL	Simulation time horizon T used as the normalisation denominator in R_D
T	REAL	Full simulation window length (seconds)
fault_t_end	REAL	Time at which the fault is cleared (seconds)
gen_detail_json	TEXT	JSON dict of per-generator angular deviation and disconnection details from the Dynawo curves
R_O	REAL	Relaxed-OPF objective value (operational severity). Set to 5000.0 if the OPF returns no solution
opf_detail_json	TEXT	JSON dict with OPF metrics: gen_dispatch_change, voltage_violations, gen_violations, branch_violations, load_shed_p/q, total_violations, objective_value, infeasible
state_features_json	TEXT	JSON dict of the post-action state feature vector (bus voltages, line loadings, etc.) computed by rl_state_features.compute_state_features after tripping all lines in previous_lines + [action_line_index]
computed_at	TEXT	ISO-8601 timestamp of when this row was first inserted

Primary key: (scenario_id, previous_lines, action_line_index)

knn_sources (auxiliary table)

Marks which transitions were collected by a purely random policy and are therefore unbiased samples suitable for KNN baseline training. A triple can be tagged without re-computing anything — the actual data lives in transitions.

Column	Type	Description
scenario_id	INTEGER	Same key as transitions
previous_lines	TEXT	Same key as transitions
action_line_index	INTEGER	Same key as transitions
run_tag	TEXT	Label for the collection run (default "random")

Primary key: (scenario_id, previous_lines, action_line_index)

---

Contents

rl_experience.db (training)

Metric	Value
Total transitions	19 801
Distinct scenarios	298
KNN-tagged transitions	3 427
Depth-0 transitions (first fault)	2 699
Depth-1 transitions (second fault)	6 382
Depth-2 transitions (third fault)	10 720

rl_experience_test.db (evaluation)

Metric	Value
Total transitions	11 469
Distinct scenarios	100
KNN-tagged transitions	0 (evaluation only — never used for KNN training)
Depth-0 transitions	1 551
Depth-1 transitions	3 539
Depth-2 transitions	5 994
Depth-3 transitions	385

---

Failure handling

When Dynawo's transient solver or pandapower's pre-fault power flow diverges, R_D is set to 5000.0 (DYNAWO_FAILURE_R_D) rather than being skipped or zeroed. Similarly, OPF failures set R_O = 5000.0 (OPF_FAILURE_R_O). Both values are deliberately above the largest organically observed rewards (R_D ≈ 1 856, R_O ≈ 80–100) so that divergence is treated as the most severe outcome, not a data gap. The gen_detail_json or opf_detail_json fields of such rows contain {"diverged": true, "error": "<message>"}.

---

Usage

from rl_experience_store import ExperienceStore

store = ExperienceStore(db_path="rl_experience.db")

# Cache-first lookup (computes and stores on a miss)
record = store.get_or_compute(
    scenario_id=5,
    net=net,                  # solved pandapower network
    previous_lines=[138],     # lines already tripped
    action_line_index=60,     # line to trip now
)

# Alpha-blend at query time (no re-simulation)
r_d, r_o, r_total = store.reward(record, alpha=0.7)

# Direct read (returns None on a miss)
record = store.get(scenario_id=5, previous_lines=[138], action_line_index=60)

---

Notes

• Both databases are append-only in normal use. INSERT OR REPLACE is used on write, so re-running a scenario with the same triple overwrites the old row.
• The rl_experience_test.db scenario IDs start at 10 000 to ensure strict separation from the training pool (IDs 0–299).
• The knn_sources table in rl_experience_test.db is empty by design — test transitions are never used to train the KNN baseline.
• State features are stored post-action (after all outages in previous_lines + [action_line_index] are applied), reflecting the network state the agent will observe at the next step.