You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Release Notes
Added
training.cut_selection.selection.lookback_iterations for level1 keeps
the trial states of only the rows added in the last lookback_iterations
iterations, which bounds their memory.
simulation.cut_set, all or active, chooses which rows of each stage
the simulation loads.
Generic-constraint expressions accept +, -, *, /, parentheses and
unary minus over numbers and @parameters, on both sides of the relation.
They must stay linear in LP variables.
A scalar parameter in constraints/generic_parameters.json may be kind: "expression", a formula over other parameters.
Stored-energy parameters in MWh: max_stored_energy_mwh, integrated_accumulated_productivity_mwh_per_hm3 and integrated_equivalent_productivity_mwh_per_hm3. Price the slack of a
constraint written with them with "penalty_basis": "per_unit".
integrated_equivalent_productivity_mw_per_m3s in system/hydro_energy_productivity.parquet sets a plant's useful-range mean
productivity for the stage outright.
A diversion channel in system/hydros.json may leave out max_flow_m3s,
or set it to null, for a channel with no capacity limit.
A resumed run restores the level1 trial states, the binding counters of max_active_per_stage and dynamic, and one solver basis per forward pass
and node. A warm-started level1 run restores the trial states, and a warm
start restores the bases of a policy trained with a selection method.
Changed
BREAKING:level1 keeps exactly the cuts that own a stored trial state:
the cut with the highest value at each state. Results change for every study
that uses level1. Retrain its policies.
BREAKING:simulation.cut_set defaults to all, so a run with level1
or max_active_per_stage simulates with every row training added,
deactivated and evicted ones included. Set it to active for the earlier
behaviour. Runs without selection or a cap, and dynamic runs, are
unchanged.
BREAKING:row_activity_tolerance defaults to 1e-6 instead of 0, so a
run with max_active_per_stage or dynamic that omits it can evict and seed
other rows. Set it to 0 for the earlier behaviour.
BREAKING:violation_tolerance of dynamic is a relative factor, absent
by default. Runs can return different cuts and bounds, and a value written
for the earlier absolute tolerance can fail at load: delete the key or raise
it.
BREAKING: a resumed run continues as the uninterrupted run would: every
cut keeps its original position, and level1 and max_active_per_stage
treat the reloaded cuts as their own. A capped resume used to count only the
cuts added after it. Under warm_start, level1 selects a loaded cut its
source scored like a generated one.
max_active_per_stage caps only the rows the current run added, not loaded
ones.
BREAKING: the cut-selection config is checked more strictly. A zero in check_frequency, lookback_iterations, start_iteration, candidate_recency or max_added_per_round, a selection given as an
array, and max_active_per_stage together with dynamic or with training.selection.method = enumerated are refused at load.
BREAKING: policy checkpoints are written with format_version 4, and
version-3 checkpoints are refused. Re-run the program that produced the
policy. Cut values are stored in the run's scaled units, so a resume with the
same modeling.cost_scale_factor is exact.
BREAKING: a run that fails writes its partial policy to <policy.path>.partial and leaves the last whole checkpoint in policy.path.
policy.mode = resume is refused, with a message naming what is missing,
when the checkpoint lacks the level1 trial states or binding counters the
run needs, holds a shorter lookback window, an incomplete basis store or
another training.selection.forward_passes, or when the rows of policy.boundary differ from its terminal pool.
A dynamic run that simulates with simulation.cut_set = all is refused
when the policy it loads holds inactive rows that all would include. Set simulation.cut_set to active or use level1.
BREAKING:training/cut_selection/iterations.parquet has one row per (iteration, pool_id), cuts_in_pool replaces cuts_populated, and pool_id and node_id join stage_id. Update readers of the file.
BREAKING:exports.states writes each pool's stored trial states at the
pool's state dimension, for every selection setting. The per-stage
visited-states export at full dimension is removed.
BREAKING: a refused policy loaded only to simulate is reported as SimulationPolicyIncompatible, by novomodelo validate --json and novomodelo.io.validate, no longer as WarmStartIncompatible. Filter on
both kinds to keep the previous behaviour.
BREAKING:system/hydro_energy_productivity.parquet must carry the integrated_equivalent_productivity_mw_per_m3s column, and equivalent_productivity_mw_per_m3s sets only the reference-point
productivity. A plant with VHA geometry now integrates its useful-range mean
productivity even with that override, so its stored-energy tags and outputs
change where the two differed. To keep a study's values, write the override
into the new column too; otherwise add the column as all-NULL.
A plant whose geometry gives a non-positive useful-range mean head is
rejected at setup even with an equivalent_productivity_mw_per_m3s override.
BREAKING: the linearized_head production model is removed: it always
ran as constant_productivity, which gives the same results. model and default_model accept only constant_productivity and fpha; any other
name used to run silently as constant productivity.
BREAKING: a positive max_diversion_m3s in constraints/hydro_bounds.parquet on a hydro with no diversion channel,
and a negative diversion.max_flow_m3s, are rejected at load. Before, the
diverted water left the system without reaching any plant.
BREAKING: an enabled generic-constraint slack must declare slack.penalty_basis: per_hour charges every hour the row covers, as
before, and per_unit charges once per unit of violation. Add "penalty_basis": "per_hour" to every enabled slack to keep a study's costs.
A disabled slack that sets penalty or penalty_basis, and a field in constraints/generic_parameters.json that does not belong to the entry's
kind, are rejected.
A generic constraint written without a relational operator drops a variable
whose coefficient is exactly zero. A row left with only stage-level terms
collapses to one stage-level row, so under per_unit a violation costs the
penalty once instead of once per block.
x86_64 Linux builds from source, cargo install included, and the MPI
release archive no longer enable AVX2 and FMA, and qhull, HiGHS and CLP
compile without FMA on every target. Results no longer depend on the build
host or the compiler's default target, and can differ in the last bits from
0.18.1's, on aarch64 too. Before, a compiler that targets FMA by default, such as the
GCC of Ubuntu 26.04, could fit an FPHA plane with a storage coefficient just
below zero, which loading the planes back as precomputed rejected.
BREAKING: the Linux CLI archives are built on Ubuntu 24.04 and need
glibc 2.38 or newer. They no longer run on Ubuntu 22.04, Debian 12 or
RHEL 9. The Python wheels keep their manylinux tags.
Removed
BREAKING: the lml1 and domination values of training.cut_selection.selection.method (use level1) and level1's tie_tolerance (delete the key). A config that names them fails at load.
Fixed
A plant's min_turbined_m3s or min_generation_mw in constraints/hydro_bounds.parquet had no effect, and its max_turbined_m3s or max_generation_mw capped each bus's share of a plant
whose unit groups sit on several buses, not the total. Both now bind on the
plant's total.
A lazy solve under dynamic that reached the round cap returned a relaxed
result.
cuts_removed in training/convergence.parquet was always 0, and cuts_active and row_pool.peak_active ignored evictions and
reactivations.
Training aborted when max_active_per_stage was below the number of cuts
loaded by policy.boundary.
A generic constraint whose terms land on the same LP column, such as a
repeated variable, aborted the run when HiGHS loaded the LP.
A checkpoint over the 2 GiB limit of the policy format, or one a load would
reject, is refused when written instead of being written unreadable.
A basis-store decode failure at the end of a multi-rank run left the other
ranks blocked.
Under MPI, the backward wait and bwd_load_imbalance_ms left out the time
a rank spends waiting for the other ranks.
novomodelo-cli 0.19.0
Install novomodelo-cli 0.19.0
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/ons-ccee-epe/novomodelo/releases/download/v0.19.0/novomodelo-cli-installer.sh | sh
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Release Notes
Changed
BREAKING: the default of modeling.inflow_non_negativity.method is now truncation_with_penalty (was penalty). Set "penalty" to keep the old
default.
BREAKING:truncation and truncation_with_penalty follow Larroyd et al.
(2022). The PAR inflow is no longer raised to zero: it stays raw, may be
negative, and carries into the next stage's lag. A per-hydro correction lifts
the inflow that enters the water balance to zero. Under truncation it equals
the shortfall and costs nothing; under truncation_with_penalty it lies
between zero and the shortfall and is priced by hydro.inflow_nonnegativity_cost in penalties.json. incremental_inflow_m3s reports the raw inflow and inflow_nonnegativity_slack_m3s the correction. Results change for studies
whose inflow draws go negative. Under either method the lower bound can exceed
the upper bound, so a gap stopping rule can report convergence too early.
Under penalty and truncation_with_penalty, the inflow slack of a plant
that has not started filling enters the water balance of the plant that
receives its inflow; it used to enter none.
Under the truncation methods, a generic constraint's hydro_inflow term
includes the correction, so it matches the inflow in the water balance.
HiGHS forward and simulation solves start cold by default; backward solves
still warm-start. Set "use_warm_start": true in training.solver.forward or simulation.solver to warm-start that phase. Retries keep the phase's
setting. A cold solve can stop at a different, equally optimal vertex, so
simulated duals and the final lower bound may differ slightly from 0.18.0.
novomodelo run writes its report (banner, progress and summaries) to stdout;
warnings and errors stay on stderr. Scripts and job logs that read the report
from stderr must read stdout. --color applies to both streams.
The training gap is printed with 5 decimals instead of 1.
Fixed
truncation no longer produces an infeasible LP when an inflow's standard
deviation is zero.
novomodelo run under MPI no longer panics and hangs when the forward passes
do not divide evenly across ranks.
A rank that panics under MPI aborts the whole job with exit code 101, instead
of leaving the other ranks waiting until the job is killed.
Under MPI, the training summary and solve_stats in training/metadata.json
no longer multiply the LP-solve counts and the forward solve time by the
number of ranks.
novomodelo-cli 0.18.1
Install novomodelo-cli 0.18.1
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/ons-ccee-epe/novomodelo/releases/download/v0.18.1/novomodelo-cli-installer.sh | sh
BREAKING: the product is renamed. The CLI is novomodelo
(novomodelo-mpi for the MPI build), the Rust crates are novomodelo-*, the
Python package is novomodelo-python (import novomodelo), and environment
variables such as COBRE_TCP_COORDINATOR are now NOVOMODELO_TCP_COORDINATOR.
Everything else behaves as in Cobre 0.18.0.
BREAKING: outputs and policy checkpoints record software as novomodelo. A policy loads only in the software that wrote it, so policies
written by Cobre are refused, and Cobre refuses this fork's.
The $schema URLs in examples, init templates and the schemas point at this
repository's schemas/ directory.
novomodelo-cli 0.18.0
Install novomodelo-cli 0.18.0
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/ons-ccee-epe/novomodelo/releases/download/v0.18.0/novomodelo-cli-installer.sh | sh
