Light environment, topological actions in ScenarioSweep, and batch slack fixes by BDonnot · Pull Request #226 · Grid2op/lightsim2grid

BDonnot · 2026-10-05T19:28:19Z

	added	removed
C++ (src/core, src/bindings)	+3338	-177
Python (package)	+191	-8
Tests	+2370	-2
Docs + changelog	+405	-0
Build / other	+294	-0
total	+6598	-187

fast-env already has dev_1.1.1 merged in (7e30c26), so this merges without conflicts.

What it brings

Light environment (LightEnv, C++ with Python bindings): a grid2op-like environment running on an LSGrid.

init_actions registers checked topological actions; step(act_id) plays them with grid2op-like cooldowns.
Protections against overflow.
LightEnvObservation: read-only views on the env's state, with no copy.
Cheap copies; noexcept moves in C++.
A docs page (docs/light_env.rst) and a benchmark against grid2op (benchmarks/light_env.py).

ScenarioSweep.set_topo_actions: one topological action per row (disconnections, reconnections, a bus split or merged), still on a single symbolic analysis.

The solver labelling is built for the union of the buses the rows use.
The admittance entries a row writes are reserved up front.
Each row is value edits only.
What is still refused is listed under [TODO] in the changelog.

Core: parameters that took bare references to Eigen objects now take Eigen::Ref.

Fixes made after merging dev_1.1.1

Each code fix came with a test that failed before it.

The slack pre-pass now counts each element where the row's action puts it (redistribute_slack + set_topo_actions). A moved generator still injects; a load or storage unit the action disconnects, or leaves alone on a busbar, is lost; a reactivated generator is a gain. A unit that can take part in the slack takes its share on its new bus.
redistribute_slack is refused with a single-slack algorithm (NRSing_*, Gauss-Seidel). Those ignore the weights, so before, the voltages came out unchanged while compute_physical_violations read targets the solve never used. New capability flag: BaseAlgo::distributes_slack().
Refactor fallback when a row has slack weights of its own. A unit saturated by the pre-pass has its weight set to 0. With KLU that zero could land on a pivot kept from the base factorization, and the row then failed to converge without handle_disconnected_grid. Reproduced on l2rpn_case14_sandbox with NR_KLU.
Topological actions on an algorithm that cannot mask a bus now get an error that names set_topo_actions (it used to name handle_disconnected_grid). Actions that change nothing are plain rows on any algorithm.
C++ tests: TempFile names carry a per-process random token. Before, ctest -j processes could write the same file: a parallel stress run failed 13 of 24 file-writing tests, and now passes.
Housekeeping: shorter changelog entries; removed a root-level scratch script.

Testing

C++ (Catch2): 398/398, serially and with ctest -j 16 --repeat until-fail:3.
Python (unittest, grid2op installed from source): test_ScenarioSweep_topology, test_batch_redistribute_slack, test_ScenarioSweep, test_ScenarioSweep_gen_contingency, test_ContingencyAnalysis_split, test_ContingencyAnalysis_limit_violations, test_physical_violations, test_lsgrid_violations, test_main_component_slack, test_can_participate_slack, test_cap_slack_at_active_limits, test_binary_serialization, test_storage_pypowsybl, test_voltage_control_batch, test_hvdc_batch, test_fdpf, test_solver_control, test_SecurityAnlysis, test_LightEnv.
The dist_slack_algorithm plugin example builds and its test_plugin.py passes.

Note for review: DCO

Six older commits on this branch (February 2025: 525803d, 6a286c5, 5f080d0, 2a34c6b, a0ca58a, 1d6c68e) have no Signed-off-by: trailer, so the DCO check is expected to fail until they are signed off.

🤖 Generated with Claude Code

https://claude.ai/code/session_01ApvmWkD2kwsiB7HscvzfuV

Generated by Claude Code

Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>

add a mode to keep only the main cc when performing steps

Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>

Sync the fast-env branch with the head of the fix_derivatives branch. The PR renamed GridModel to LSGrid, moved the core under src/core and split the pybind11 module into src/bindings/python/binding_*.cpp, so the light-env work carried on this branch is ported onto that layout: - src/light_env/*.hpp move to src/core/light_env/, are wrapped in the ls2g namespace, and target LSGrid (change_bus_* now takes a GridModelBusId). - the LightEnv / Protections / TopoAction / ElementType bindings that lived at the end of src/main.cpp become src/bindings/python/ binding_light_env.cpp, registered through bind_light_env(). - test_light_env.py imports from lightsim2grid.lightsim2grid_cpp. The three modify/delete conflicts (src/GridModel.hpp, src/element_container/SGenContainer.cpp, src/main.cpp) are resolved by taking the deletion: the fast-env edits there were a template parameter rename, a blank line, and the bindings ported above. Assisted-by: Claude Code (claude-fable-5-1) Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>

…cooldowns LightEnv.init_actions (lightsim2grid/lightEnv.py) takes a list of grid2op actions (set_bus / set_line_status only) or TopoAction objects. Every action is checked in C++ against the initial grid: a non-existing element, a busbar outside -1 / 0 / 1..n_busbar_per_sub (eg -2), a bad status or a contradiction between set_line_status and set_bus on a line end raises a ValueError naming the action, and nothing is registered. TopoAction now speaks grid2op: local busbar ids resolved through the substation of each element (new get_subid getters on the containers), grid2op line numbering (powerlines then transformers), set_line_status. step(act_id) applies the action when legal. Cooldowns follow grid2op's rules and update order: an action touching a substation / line still in cooldown is illegal (info["is_illegal"]), elements the action touched get nb_timestep_cooldown_sub / nb_timestep_cooldown_line, lines tripped by the protections get nb_timestep_reconnection. reset restores the initial topology and the protections' counters. Two pre-existing bugs fixed on the way: the protections looped for ever once a line had tripped (its overflow counter never went down), and the observation returned by reset was a stale rho. Tests (lightsim2grid/tests/test_LightEnv.py) compare every step with a grid2op env on the same grid and chronics: illegal flag, cooldown vectors, line statuses, load / gen buses and flows. Assisted-by: Claude Code (claude-fable-5-1) Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>

ScenarioSweep.set_topo_actions takes one TopoAction per row (grid2op semantics: set_bus / set_line_status; the grid2op wrapper converts grid2op actions), checked against the grid when registered. This first stage plays disconnections only, through the same fixed-sparsity paths as the masks, so the whole sweep still runs on one symbolic analysis: - a branch (set_line_status -1, set_bus -1 on an end) joins the row's disconnected-branch list (YbusPolicy::Contingency::topo_branches_off, merged by branch_ids_for_row: same Ybus edit, connectivity, current checks and flow cleaning as a mask entry); - a generator (set_bus -1) joins the generator-contingency path (SbusPolicy::Vary::topo_gen_off, read with gen_off through gen_off_in: injection, slack re-weighting, PV -> PQ pinning); - a load or a storage unit (set_bus -1) is taken back out of the row's injection (topo_loads_off / topo_storages_off). A row naming an element both in a mask and in its action, a move (set_bus > 0), a reconnection and the DC algorithm are refused with a message; the changelog TODO section lists them. TopoPlan.hpp resolves an action into per-element placements against the base grid (the representation the next stages extend). TopoAction exposes its resolved entries and is bound with apply_to_gridmodel. Tests compare every row with a one-off ac_pf on a copy of the grid with the action really applied (voltages and flows), and pin nb_analyze == 1, threads agreement, the masked / NOT_SIMULATED split-row behaviour and the refusals: lightsim2grid/tests/test_ScenarioSweep_topology.py and src/tests/test_scenario_sweep_topology.cpp. Assisted-by: Claude Code (claude-fable-5-1) Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>

…PQ -> PV) A row may put back a generator disconnected in the base grid, on the bus it was last on (set_bus to that busbar). Its bus turns PV for the row at constant sparsity: a base-PQ bus already owns a Vm unknown and a Q equation, so the row only pins that equation (set_pv_pinned_buses) and seeds |V| there at the generator's set-point -- the row's own when modify_gen_v gives one. A generator that does not regulate voltage is an injection and its bus keeps its label. - _maybe_prepare_gen_contingency now builds a per-row pinned set (_row_pv_pinned_: the switchable buses still PV plus the buses a reactivated generator pins) and per-row |V| seeds (_row_vm_seed_, applied by _apply_step_topo_seed after _apply_step_gen_v). A base-PV bus whose own controllers the row takes out stays PV through a reactivated generator, provided both ask the same magnitude. - SbusPolicy::Vary::topo_gens_on adds the generator's P (and Q if it does not regulate) to the row; VoltageSourceContainer gains would_be_local_voltage_controller (the status-free predicate). - refused with a message: a slack participant, a slack bus, a remote or group-held controller, another busbar, keep_jacobian / compute_physical_violations with a reactivation. Tests: the stage 1 file runs again on a base grid with a generator off (TestScenarioSweepGenReactivation) plus the reactivation rows against a one-off ac_pf, with and without a per-row set-point, the refusals, and nb_analyze == 1; the Catch2 test adds a non-regulating base-off generator and the same cases. Assisted-by: Claude Code (claude-fable-5-1) Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>

…, reconnections) A row may now move elements between busbars and reconnect branches -- what creates a bus (a busbar nothing stood on, now used) or merges two -- with the whole sweep still on one symbolic analysis. The union layout: - LSGrid::init_converter_bus_id / build_solver_input take an optional list of buses to keep in the solved system although empty in the grid; the sweep hands it every bus any row uses (_maybe_topo_union, L1). - the admittance entries a row writes (the diagonal of every extra bus, the four entries of every branch placement) are reserved as stored zeros right after the cache is built (_maybe_reserve_union_pattern), so the base solve builds the superset Jacobian sparsity; extra buses are seeded with a finite |V| and masked (identity rows) in the base case and in every row that leaves them empty (_base_masked_, the masked loop's resting state -- the masked loop runs whenever actions are registered). - a row is value edits: YbusPolicy::Contingency::topo_branches_moved (the base contribution out, the raw block in at the row's buses, negated coefficients through the existing -= path); SbusPolicy::Vary::topo_loads_on / topo_storages_on / topo_gens_on (with their gridmodel bus); the generator pinning of stage 2 (a bus a generator lands on is held, a bus its last generator leaves is solved for); per-row connectivity by the search whenever a row places a branch, with connected meaning nothing stranded beyond the base mask, and NOT_SIMULATED kept without handle_disconnected_grid. - flows and current-limit checks read a row's own branch placements (BranchBusOverride, BaseBatchSolverSynch::_row_branch_overrides_). Refused: moving or reactivating a slack participant, a generator on a slack bus or one regulating remotely / in a control group, a regulating storage unit, keep_jacobian / compute_physical_violations with a generator move, the DC algorithm (changelog TODO). Tests: substation splits and merges, a dangling bus, a generator moved to a new busbar, an island of one bus (masked), reconnections where the branch was and on a new busbar, two substations rewired in one row, flows and current violations of moved lines, four threads vs one, and nb_analyze == 1 throughout -- every row against a one-off ac_pf on the rewired grid (lightsim2grid/tests/test_ScenarioSweep_topology.py, src/tests/test_scenario_sweep_topology.cpp). Assisted-by: Claude Code (claude-fable-5-1) Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>

Assisted-by: Claude Code (Fable 5.1) Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>

Follows 57e73e1 (compute_physical_violations following a row's generator placements), whose Python tests did not run: - the pandapower loader expects every busbar to exist in the network (n_sub x n_busbar buses): the test adds busbar 2 of each substation before loading case14 with n_busbar_per_sub=2; - the reference rebuilt from the base grid instead of the grid under test, so the reactivation case compared with the generator on; - the "every overloaded bus is reported" check now skips the slack bus, which the reactive check does not cover on a plain row either. Assisted-by: Claude Code (claude-opus-5) Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>

step() returned current_step_ / max_iter_ as reward: an integer division, and by the powerflow iteration cap rather than the episode length, so the reward was 0 for most of an episode. info["survival_time"] was current_step_ / max_step_, also an integer division, so always "0" or "1". Both are now survival_ratio(), the fraction of the episode survived as a double (1 on success, current_step / max_step otherwise). The bindings of LightEnv.reset / assign_time_series and of the thermal limit and max overflow getters / setters of Protections had "TODO" as docstring; they are documented (units, numbering, which injections are actually applied), and the arguments are named. Tests: test_LightEnv (reward over a short episode, survival_time on success and on a divergence). Assisted-by: Claude Code (claude-opus-5-5) Claude-Session: https://claude.ai/code/session_019LN3o2bJBuNuNf9sN2tcNb Signed-off-by: Benjamin Donnot <benjamin.donnot@rte-france.com>

reset() and step() returned a copy of rho. They now return the env's LightEnvObservation, which holds only a pointer to its env; every getter returns an Eigen::Ref<const ...> and the bindings expose them as read-only numpy views (def_property_readonly, so reference_internal: the array keeps the observation alive, which keeps the env alive). What it exposes: rho, p / q / a on both sides of every line (grid2op numbering, powerlines then trafos, currents in kA like the thermal limits), topo_vect (local busbar ids at their grid2op position), the line and substation cooldowns, load_p, gen_p and current_step. Where the data lives: - rho, load_p, gen_p, the cooldowns: read where they already are (the protections, the grid results, the env), nothing is copied. - the flows and topo_vect are not stored in grid2op order anywhere (lines and trafos are separate containers, the grid knows buses, not local busbars), so extract_observation writes them once per step into env-owned buffers, which the observation then views. Lifetime: the env buffers (flows, topo_vect, cooldowns) are allocated once in the constructor and only written in place (reset_cooldowns now setZero()s instead of reallocating), so a view on them stays valid for the life of the env. rho and the grid results are reallocated at reset (new LSGrid, Protections::check_validity) and on a divergence (reset_results), so a view is documented as valid until the next reset or a step ending the episode by a divergence. LightEnv is made non-copyable and non-movable since its observation points to it. topo_vect needs the grid2op positions (set_*_pos_topo_vect, done by LightSimBackend): new const getters get_pos_topo_vect[_side_1/2] on the containers, and SubstationContainer::gridmodel_to_local, the inverse of local_to_gridmodel. Without the positions, topo_vect raises. Tests: test_LightEnv. assert_same_state now also compares the light observation with the grid2op one at every step of the existing tests (topo_vect, flows, load_p, gen_p, cooldowns); new tests check the arrays are read-only views that follow the env across a step, and that a view keeps its env alive. Assisted-by: Claude Code (claude-opus-5-5) Claude-Session: https://claude.ai/code/session_019LN3o2bJBuNuNf9sN2tcNb Signed-off-by: Benjamin Donnot <benjamin.donnot@rte-france.com>

A LightEnv could not be copied: its live grid is a unique_ptr (LSGrid has a copy constructor but no copy assignment) and its observation points back to it. A copy is what a tree search (or any look-ahead) on the env needs. The state now lives in a LightEnvState base, copied member by member, and LightEnv only adds the observation: its copy / move constructors and assignments copy the base and keep the observation pointing to the env it belongs to. Having the state apart means a member added later cannot be forgotten by a hand-written copy. The live grid is held by a small DeepCopyPtr (a unique_ptr whose copy clones). What never changes during an episode is shared between copies through a shared_ptr<const ...>: the initial grid, the time series (now one TimeSeries struct) and the registered actions. assign_time_series / init_actions replace that env's pointer only, so the copies do not see each other's changes. Everything else (live grid, protections and their counters, cooldowns, step, observation buffers) is copied: a copy is an independent env at the same point of the same episode. On case14 a copy takes ~11 us, a step ~17 us. reset() without assign_time_series now raises instead of reading row 0 of an empty matrix. Python: LightEnv(other), __copy__ / __deepcopy__ in the bindings, and LightEnv.copy() / __copy__ / __deepcopy__ in lightEnv.py so a copy keeps the python subclass (its init_actions accepting grid2op actions). Tests: test_LightEnv (a copy has the same state in its own memory, steps without touching the original and gives the same result on the same step, outlives the original, and does not share a later assign_time_series). Assisted-by: Claude Code (claude-opus-5-5) Claude-Session: https://claude.ai/code/session_019LN3o2bJBuNuNf9sN2tcNb Signed-off-by: Benjamin Donnot <benjamin.donnot@rte-france.com>

The move constructor / assignment added with the copies were not noexcept, so a std::vector<LightEnv> copied every env when it grew instead of moving them. Every member of LightEnvState is nothrow movable (LSGrid itself is not, it is behind DeepCopyPtr), so the moves are now noexcept; two static_asserts next to LightEnvState keep it that way when a member is added. A move copies nothing: the grid, the protections and the observation buffers change owner, and the observation of the env moved to views them at the same address. A moved-from env had a null grid: get_grid(), reset(), step() or obs.load_p / gen_p dereferenced it. It can now only be destroyed or assigned to (which makes it usable again); anything else throws std::logic_error naming the call. step() also checks, so it does not ask for a reset() that would throw in turn. Tests: new src/tests/test_light_env_copy_move.cpp (Catch2), since the moves only exist on the C++ side: copies are independent and step like the original; a move keeps the buffers and the grid at their address and steps like the original would have; a moved-from env throws and is usable again once assigned to; envs in a growing std::vector keep their own observation. Full C++ suite green (353), and the [light_env] cases under the ASan + UBSan build (__SANITIZE=1). Assisted-by: Claude Code (claude-opus-5-5) Claude-Session: https://claude.ai/code/session_019LN3o2bJBuNuNf9sN2tcNb Signed-off-by: Benjamin Donnot <benjamin.donnot@rte-france.com>

The Windows wheels failed on the static_assert guarding the noexcept moves: light_env.hpp(195,66): error C2338: static_assert failed: 'LightEnvState should be nothrow move constructible' MSVC's std::unordered_map allocates a sentinel node, so its move constructor is not noexcept (libstdc++'s is, hence green on Linux and macOS), and LightEnvState held one: the step info. That map was never state: reset() and step() cleared it first thing, filled it and returned a copy of it. It is now a local of step() (reset() returns an empty one), which removes the member and a map copy per step. Every remaining member is a scalar, an Eigen vector, a std::vector, a shared_ptr or a unique_ptr, all nothrow movable on MSVC too. The comment on the static_asserts now names the containers to keep out of the state. Checked: the three failed Windows jobs all stop on that one assert and nothing else; C++ [light_env] cases and test_LightEnv green on Linux. Assisted-by: Claude Code (claude-opus-5-5) Claude-Session: https://claude.ai/code/session_019LN3o2bJBuNuNf9sN2tcNb Signed-off-by: Benjamin Donnot <benjamin.donnot@rte-france.com>

light env: zero-copy observation, copyable / movable LightEnv, reward fix

Recent commits (light env, ScenarioSweep topology, slack participation, voltage-control override) introduced parameters taking bare `const RealVect &` / `RealMat &` / `CplxVect &` where the rest of the code base uses Eigen::Ref. Aligned them with the convention: - read-only inputs -> `const Eigen::Ref<const T> &`: InjAction ctor, LightEnv::assign_time_series (bound to Python), LightEnv::aux_fill_topo_vect, TopoAction::aux_check_subid / aux_resolve_one_side, and the set_voltage_control_v_set chain (BaseAlgo, NRAlgo, NRSystem, AlgorithmSelector, VoltageControl::set_v_set_override); - in-place outputs that are never resized -> `Eigen::Ref<T>`: SlackParticipation::accumulate_raw / split, the generator and storage accumulate_slack_weights_solver, BaseBatchSweep::_maybe_seed_extra_buses and _apply_step_topo_seed. Left as they are on purpose: getters returning a const reference to a member (get_subid, get_pos_topo_vect...), local aliases of members, and output buffers that are resized or returned by reference. No behaviour change. Assisted-by: Claude Code Claude-Session: https://claude.ai/code/session_01Ne9bEVwN1gxzvfKmccCj3g Signed-off-by: Benjamin Donnot <benjamin.donnot@rte-france.com>

docs/light_env.rst: what LightEnv is for, a quick start building it from a grid2op env, how to set up the time series / protections / cooldowns / actions, what reset and step return, the observation (a zero-copy view), copying for lookahead, the list of what it does not model compared with grid2op, and the benchmark results. benchmarks/light_env.py: the same l2rpn_case14_sandbox chronics played on a grid2op env with LightSimBackend and on a LightEnv built with the same thermal limits, protections and cooldowns (grid2op's hard overflow threshold raised out of reach, the light env has none). Three workloads: do nothing, a random unitary set_bus action every 10 steps, and a one-step lookahead over 10 candidates (obs.simulate vs light_env.copy().step). Both sides play the same number of steps and end the same episodes. On the test chronics a step is 25 to 50 times faster than grid2op's, a lookahead evaluation about 20 times faster than obs.simulate. The first step of a copy is ~3x slower than a warm step: a copied LSGrid starts with a cold solver. Assisted-by: Claude Code Claude-Session: https://claude.ai/code/session_01Ne9bEVwN1gxzvfKmccCj3g Signed-off-by: Benjamin Donnot <benjamin.donnot@rte-france.com>

set_thermal_limit_or / _ex and set_max_line_time_step_overflow took a mutable Eigen::Ref, which pybind11 can only bind to a writeable array of the exact dtype: grid2op's thermal limits (float32) were refused, and so was any read-only array. They only read their input, so they now take `const Eigen::Ref<const T> &`, the convention elsewhere in the core. Found while checking the snippets of the new documentation page. Assisted-by: Claude Code Claude-Session: https://claude.ai/code/session_01Ne9bEVwN1gxzvfKmccCj3g Signed-off-by: Benjamin Donnot <benjamin.donnot@rte-france.com>

Eigen::Ref convention in recent core code; light env docs page and benchmark

Conflicts resolved: - BaseBatchSweep.hpp: dev_1.1.1 moved the operational limit checks to OperationalCheck.hpp (and added violation_rel_tol); the per-row branch placement fast-env gave check_current_violations (BranchBusOverride, a branch a topological action moved read with its raw admittance block) is ported there. BranchBusOverride moves with it, so that LSGrid.cpp can keep including OperationalCheck.hpp without the batch headers. _select_ref_slack_and_masks keeps dev_1.1.1's solver-numbered reference slack and fast-env's "no handle_disconnected_grid: a row stranding a base bus stays NOT_SIMULATED" line; the call sites pass both rel_tol and the row's overrides; the topology and slack-redistribution members and helpers are both kept. - StorageContainer.hpp: dev_1.1.1's storage_off mask on accumulate_slack_weights_solver, with fast-env's Eigen::Ref output. - CHANGELOG.rst: dev_1.1.1's rewritten file, with fast-env's unreleased entries moved under [1.1.1] and its TODO entry kept. Assisted-by: Claude Code Claude-Session: https://claude.ai/code/session_01ApvmWkD2kwsiB7HscvzfuV Signed-off-by: Benjamin Donnot <benjamin.donnot@rte-france.com>

With redistribute_slack, _prepare_slack_redistribution works out what a row loses before the solve and shares it on the slack units within their limits. It read every element off the base grid, so a row carrying a topological action (set_topo_actions) was mis-counted: - a generator the action moves is in topo_gen_off, so it was counted as lost power although it still injects on its new bus; - a load or storage unit the action disconnects, or leaves alone on a busbar (masked), was not counted at all, nor a generator it reactivates; - the elements moved off a busbar the row leaves empty were counted as stranded on that (masked) bus; - a unit only flagged "can participate in the slack" was given its share on its base bus (storage) or left out (generator) when the action moved it. The pre-pass now adds a term (d): for every unit the action places, disconnects or reactivates, what the base balance counted for it minus what the row's main component gets from it. Terms (a) and (b) leave those units out, and the participants the action places stand on the bus it gives them. A row without an action is computed exactly as before. Tested against a one-off grid with the action applied and the true lost power redistributed (test_ScenarioSweep_topology, the RedistributeSlack classes): every row but the plain one and the generator disconnection failed before. Assisted-by: Claude Code Claude-Session: https://claude.ai/code/session_01ApvmWkD2kwsiB7HscvzfuV Signed-off-by: Benjamin Donnot <benjamin.donnot@rte-france.com>

…ithm A single-slack algorithm (NRSing_*, Gauss-Seidel) ignores the slack weights and leaves the whole imbalance on the reference bus. With redistribute_slack the pre-pass still moved the set-points by those weights and took the units it saturated out of them: the solve did not use any of it, the voltages came out bit for bit as without the option, but compute_physical_violations read the pre-pass targets -- a generator pushed far below its min_p was no longer reported. This is the default algorithm of a grid2op environment's grid. compute() now refuses the combination. The capability is a new BaseAlgo flag, distributes_slack(), following the existing ones: true for the NR with the MultiSlack extension (a compile-time trait on NRSystem), the DC and the fast-decoupled algorithms; false for NRSing_*, Gauss-Seidel and, by default, a plugin -- the dist_slack_algorithm example declares it. Assisted-by: Claude Code Claude-Session: https://claude.ai/code/session_01ApvmWkD2kwsiB7HscvzfuV Signed-off-by: Benjamin Donnot <benjamin.donnot@rte-france.com>

Registered topological actions turn the masked row loop on (_mask_mode) without handle_disconnected_grid. On an algorithm that cannot mask a bus (Gauss-Seidel, the fast-decoupled one), compute() then failed with "the `handle_disconnected_grid` mode requires a Newton-Raphson algorithm", an option the caller never set. The error now says that set_topo_actions needs a Newton-Raphson algorithm, and names the active one. Assisted-by: Claude Code Claude-Session: https://claude.ai/code/session_01ApvmWkD2kwsiB7HscvzfuV Signed-off-by: Benjamin Donnot <benjamin.donnot@rte-france.com>

… algorithm _mask_mode() turned the masked row loop on as soon as topological actions were registered, even when every one of them changes nothing. On an algorithm that cannot mask a bus (Gauss-Seidel, the fast-decoupled one) such a batch was then refused, although nothing is ever masked -- while the DC test already pins "a do-nothing action is a plain row". It now keys on _topology_active_ (some row's action does something), which _maybe_resolve_topology settles before any caller reads it. Assisted-by: Claude Code Claude-Session: https://claude.ai/code/session_01ApvmWkD2kwsiB7HscvzfuV Signed-off-by: Benjamin Donnot <benjamin.donnot@rte-france.com>

A development script (it loads l2rpn_idf_2023 and times a LightEnv), not a unittest and referenced nowhere. benchmarks/light_env.py and docs/light_env.rst now cover what it was for. Assisted-by: Claude Code Claude-Session: https://claude.ai/code/session_01ApvmWkD2kwsiB7HscvzfuV Signed-off-by: Benjamin Donnot <benjamin.donnot@rte-france.com>

The fast-env entries under [1.1.1] and the set_topo_actions [TODO] item ran to six to thirteen lines each; brought to the one-to-four-line register CLAUDE.md asks for. The detail lives in the docs and the commit messages. Assisted-by: Claude Code Claude-Session: https://claude.ai/code/session_01ApvmWkD2kwsiB7HscvzfuV Signed-off-by: Benjamin Donnot <benjamin.donnot@rte-france.com>

…ts own A row whose slack pre-pass saturates a slack unit (redistribute_slack) solves with that unit's weight at 0. The weight is a value of the Jacobian's slack_absorbed column, which the base ("n") factorization had non-zero, and KLU's refactorization keeps the base pivot sequence: when the pivot sits on that entry, klu_refactor meets a zero and the row fails. With handle_disconnected_grid the rows took the masked path, which already turns the fallback on, so the failure only showed without it. Observed on l2rpn_case14_sandbox with NR_KLU, generator 1 joining the slack of generator 5 with a tight range: every row saturating it failed, any load change. SparseLU (it re-pivots) and NRRefactorRetry_KLU (one extra factorize) solved the same rows. Whether a unit hits the pivot depends on KLU's ordering of the grid: generator 0 in its place does not. dev_1.1.1's tests clamp a unit that does not. compute() now turns set_refactor_fallback on, for the member algorithm and the threads' ones, whenever some row solves with slack weights of its own (a participant its pre-pass saturates, or one it disconnects -- the same zero, not reproduced there), as it already does for masked buses and PV / PQ switches. The fallback only runs on a failure. Assisted-by: Claude Code Claude-Session: https://claude.ai/code/session_01ApvmWkD2kwsiB7HscvzfuV Signed-off-by: Benjamin Donnot <benjamin.donnot@rte-france.com>

ctest runs every TEST_CASE in a process of its own, and with `ctest -j` they run side by side in the build directory. TempFile named its files from a per-process counter only, so the first file of two concurrent processes was ls2g_unit_test_0.lsb in both, and one test read or removed the other's file. Reproduced by running the 24 tests that write files with `ctest -j 16 --repeat until-fail:30`: 13 of them failed. With a 64-bit token drawn once per process in the name (std::random_device mixed with the clock and the address of a static, plain C++14, no platform #ifdef) all 24 pass the same run, and the whole suite passes `ctest -j 16 --repeat until-fail:3`. Assisted-by: Claude Code Claude-Session: https://claude.ai/code/session_01ApvmWkD2kwsiB7HscvzfuV Signed-off-by: Benjamin Donnot <benjamin.donnot@rte-france.com>

…e.com> I, DONNOT Benjamin <benjamin.donnot@rte-france.com>, hereby add my Signed-off-by to this commit: 525803d I, DONNOT Benjamin <benjamin.donnot@rte-france.com>, hereby add my Signed-off-by to this commit: 6a286c5 I, DONNOT Benjamin <benjamin.donnot@rte-france.com>, hereby add my Signed-off-by to this commit: 5f080d0 I, DONNOT Benjamin <benjamin.donnot@rte-france.com>, hereby add my Signed-off-by to this commit: 2a34c6b I, DONNOT Benjamin <benjamin.donnot@rte-france.com>, hereby add my Signed-off-by to this commit: a0ca58a I, DONNOT Benjamin <benjamin.donnot@rte-france.com>, hereby add my Signed-off-by to this commit: 1d6c68e Assisted-by: Claude Code Claude-Session: https://claude.ai/code/session_01ApvmWkD2kwsiB7HscvzfuV Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>

BDonnot and others added 30 commits December 18, 2024 15:54

test to fix issue with disconnected slack [skip ci]

ec0c082

Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>

trying to debug slack disconnection bug [skip ci]

db186eb

Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>

in the middle of fixing the bug with the disconnected gen [skip ci]

cf80467

Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>

in the middle of refacto cpp side [skip ci]

b5c24c4

Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>

Merge branch 'master' into bd-dev

13e7b37

adding some links for other powerflow analysis tools [skip ci]

38e1e3f

Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>

synch with master [skip ci]

80013b8

it compiles after refacto, need to test before continuing

1d6c68e

fix an issue with shunts

a0ca58a

making sure #101 is fixed

2a34c6b

prepare for the detachment feature [skip ci]

5f080d0

fixing broken tests

6a286c5

fix broken tests before continuing the refacto

d8cb95f

Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>

fix and improve tests

0d26c76

Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>

fixing a bug in storage units

00af0eb

Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>

fix CI issue for grid2op compat

0b30b1f

Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>

add a mode to keep only the main cc when performing steps

aad6c47

Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>

Merge pull request #107 from Grid2op/bd-dev

f0f7a3f

add a mode to keep only the main cc when performing steps

update changelog and version [skip ci]

7b31c39

Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>

start the feature [skpi ci]

525803d

almost ready for the first alpha version of the light env [skip ci]

f5238a9

Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>

it appears to compile ok [skip ci]

58557fc

Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>

still bit of work on the light env

8df8429

Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>

forget to add a file for the light environment

c35c5d4

Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>

synch with version 0.12.0

5504e80

Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>

BDonnot and others added 22 commits September 18, 2026 18:08

WIP: start to add actions in batch [skip ci]

57e73e1

Assisted-by: Claude Code (Fable 5.1) Signed-off-by: DONNOT Benjamin <benjamin.donnot@rte-france.com>

Merge pull request #214 from Grid2op/claude/python-api-exposure-rwjrdx

b04dc31

light env: zero-copy observation, copyable / movable LightEnv, reward fix

Merge pull request #217 from Grid2op/claude/eigen-ref-syntax-dhrwza

72e1f4c

Eigen::Ref convention in recent core code; light env docs page and benchmark

BDonnot mentioned this pull request Oct 8, 2026

Fixes from the review of #226: topological actions in ScenarioSweep, LightEnv #228

Open

Navigation Menu

Repository navigation

Light environment, topological actions in ScenarioSweep, and batch slack fixes - #226

Light environment, topological actions in ScenarioSweep, and batch slack fixes#226
BDonnot wants to merge 52 commits into
dev_1.1.1from
fast-env

BDonnot commented Oct 5, 2026

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

BDonnot commented Oct 5, 2026

What it brings

Fixes made after merging dev_1.1.1

Testing

Note for review: DCO

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant