CHANGELOG.md

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Unreleased

Added

• investmentblock_solver solves an InvestmentBlock whose inner Block is an SDDPBlock, from a Block file as from a problem file: the UCBlock of each stage, which sits behind a BendersBFunction that no BlockConfig crosses, gets the -B "meta"-BlockConfig, and each SDDPGreedySolver passes the final state of a stage to the next; the problem file used to leave the SDDPBlock with no Solver, and the Block file refused it. examples/instance-3 is such an instance, with 3 stages and 5 scenarios, solved with -S BSPar.txt -B SDDPBCfg-LD.txt; with config/SDDPBSCfg.txt in IFCfg-SDDP.txt the stages are solved via LagrangianDualSolver, and the components that must be hard and whose primal solution the SDDPSolver needs are named by class in BSPar-LD.txt

• svm_solver estimates the leave-p-out error on a sample of the subsets (-P), repeats an estimate with one seed per repetition (-r), writes the seconds and the score of each point of the grid on each split (--csv), asks for the bias regularised with the weights (-R), eliminates the features recursively with each round optionally reoptimized (-F), walks the grid along C keeping the Solver and reoptimizing from the previous point (-W), removes each fold from one model and puts it back, the Solver re-optimizing after each change (-u), and prints the wall-clock time of the whole selection and the sum of the times of its trainings; it starts no more workers than there are trainings to run

• svm_solver/modelsel, the computational study on the model selection of an SVM: the configuration of the Solvers compared, the download of the data sets, the campaign guarded by the load of the machine and the tables

• tssb_solver/ldld, the computational study on the nested and the recursive Lagrangian dual of a TwoStageStochasticBlock: a driver of its own (ldld_bench, which runs one Solver of a BlockSolverConfig per process and prints its bound, time, iterations and peak memory, with an optional primal recovery), the configuration of the six methods, the generation of the instances, the campaign guarded by the load of the machine and the figures and tables; it depends on nothing else in the tools and is built on its own, with CMake against an installation or with its makefile

• -v 2 prints the parameters of the Solver attached to the Block and -v 3 those of the Solver of the sub-Block as well, which is how a run says what it was actually asked, rather than what the configuration files seem to say

• ucblock_solver/lagrangian_timing/rules_cost.py, which says what the operating rules of a nuclear unit cost inside a Lagrangian decomposition: it reads the CSV that the timing harness writes over two dump sets of the same fleet, the same demand and the same regime, one carrying the whole set of the rules and one reduced to the original model, and reports, per Solver, the median and the worst solve, how many ended on the time limit rather than on the problem, and the factor between the two arms

• tssb_solver also solves a MultiStageStochasticBlock when the module is built, -k giving the Benders form over the leaves of its scenario tree

• -k, --benders in tssb_solver: the Benders form of the problem is assembled around it [see TwoStageStochasticBlock::get_Benders_form()] and the BlockSolverConfig of -S is applied to its root, which is where a BendersDecompositionSolver is attached; the tool links the BendersDecompositionSolver when CMake finds it, the makefile does not

Changed

• cflblock_solver links LagrangianDualSolver, BundleSolver and BranchAndXSolver when they are built, as ucblock_solver does, so that a CapacitatedFacilityLocationBlock can be solved by Lagrangian decomposition and by a Branch-and-Bound on it

• a meta BlockSolverConfig is applied father-first, as the meta BlockConfig and the other dispatches by classname already were, and the sub-Block of a Block are looked up after it has been configured

• the hard components of the Lagrangian Dual in BSPar-LD.txt and BSPar-greedy-LD.txt of sddp_solver, and in BSPar-LD.txt of investmentblock_solver, are named with vstrNoEasy of the inner BundleSolver, which replaces vstr_LDSl_NoEasy of LagrangianDualSolver: which components are easy is a concept of BundleSolver, and the classes listed are the same

• the components whose primal solution the SDDPSolver and the SDDPGreedySolver need are named by class in the Configuration of get_var_solution() of their extra Configuration (SDDPSCfg-LD.txt, SDDPSCfg-greedy-LD.txt of sddp_solver, SDDPCfg.txt of investmentblock_solver), a SimpleConfiguration< std::map< std::string , Configuration * > > that LagrangianDualSolver reads, rather than with vstr_LDSl_VarSol in the BlockSolverConfig of the stages, which LagrangianDualSolver no longer has; the duals of the stages of SDDPSCfg-LD.txt, which were named with vstr_LDSl_DualSol, are those of the HydroSystemUnitBlock and of the relaxed constraints in the Configuration of get_dual_solution() of the SDDPSolver, which the BendersBFunction of each stage passes on

• the tools no longer link with --no-as-needed when the libraries are shared, since every module asks the linker itself to keep it (on ELF and on macOS, with MSVC by the symbol that forces it in); only the archives of a static build are still taken whole

• sddp_solver builds no configuration in code: -B is a "meta" BlockConfig dispatched to the Blocks of the inner Block of each stage (SDDPBCfg.txt by default, an ordinary BlockConfig is still applied to the SDDPBlock), and the stages are solved via LagrangianDualSolver with -B SDDPBCfg-LD.txt -S SDDPSCfg-LD.txt, or -s -B SDDPBCfg-LD.txt -S SDDPSCfg-greedy-LD.txt: the components that are hard and those whose primal and dual solution are needed are listed by class in BSPar-LD.txt and BSPar-greedy-LD.txt (vstrNoEasy of the inner BundleSolver) and in the Configurations of get_var_solution() and get_dual_solution() of SDDPSCfg-LD.txt and SDDPSCfg-greedy-LD.txt, so that they hold for any instance; with no -S the tool stops rather than building a default one

• everything investmentblock_solver gives the InvestmentBlock and its InvestmentFunction comes from the configuration files, the tool only reading and applying them: the OBlockConfig of the InvestmentBlock (config/IBOCfg.txt, the InvestmentBlock entry of the "meta"-BlockConfig InnerBCfg.txt) reformulates the bounds on the investment and gives the InvestmentFunction its ComputeConfig (config/IFCfg.txt), with the file of the investment candidates and, in the extra Configuration, the BlockSolverConfig of the inner Block; the BlockSolverConfig of the BundleSolver no longer carries strInnerBSC, a parameter that was not one and that the tool took out before applying it, nor does the tool fall back to BSCfg.txt, set the file of the candidates, reformulate the bounds or build a default formulation of the inner Block in code. A Block file now needs a BlockConfig (-B, by default InnerBCfg.txt)

• the makefile asks for -O3 -DNDEBUG and nothing else, the macro of the patch for boost::any on macOS having no reason to be there since there is no boost::any left in the core

• the set of instances is named pypsa-data, as the folder that holds it

• the parameter that the feasibility cut of the Benders decomposition needs is in the configuration, commented where the Solver refuses it, so that a run that wants that cut is one line away instead of a search through the documentation

• the master of the Lagrangian dual asks Gurobi for its least numerical care and not for none of it, satisfies its own rows tighter than the oracle satisfies its own, and declares the residual zero on the scale of the model: the extra care costs at every one of the thousands of solves of a run, while what the master needs is to be solved consistently

• the configurations of the tools move to the parameter set of BundleSolver 2.0, the Solver of the master being configured where the master is and not where the bundle is, and Method and NumericFocus counting among the integer parameters of that Solver

• a solve that the license service refuses is waited out and tried again, and a sweep that stops part way through says so and leaves no file that looks like a measurement: the units of a fleet differ from one another, so the fleet is timed whole or not at all

• the timing study reaches the nuclear units, and the MILP it compares against separates the Perspective Cuts, so that the two arms are the same model solved in two ways

Removed

• the options -R and -j of smspp_tssb_solver: a feasible solution out of a Solver that gives a bound comes from the configuration, as everything else a Solver does, i.e., from a PrimalProximalHeur (strRecoveryBSC) on the extensive form and from a BendersDecompositionSolver (strRecoveryBSC) on the Benders form, which report it as their upper bound

• the options -l, -n, -r and -s of investmentblock_solver, which had no effect: the cuts that -l named were never loaded, the number of sub-Blocks per stage of -n was never set, -r only relaxed the default formulation built in code, which is gone, and -s was read and never used; -n is again the standard option

• the options -r and -z of sddp_solver: -r only said that it no longer existed, and the hydro system as an easy component of -z is a LagrangianDualSolver configuration (vstrNoEasy without HydroSystemUnitBlock, vstr_LDSl_Cfg and int_InnerS_WVarSCfg), as are config/get_var_solution_bundle.txt and config/BendersBSCfg.txt, which only the code building the configuration read

Fixed

• cleanup_bsc() also cleans the Block configured by the "*" entry of a meta BlockSolverConfig, whose Solver were never removed

• tssb_solver -k writes (-O) the Solution of the TwoStageStochasticBlock, one ScenarioSolution per scenario as without -k, rather than that of the root of its Benders form, which pypsa2smspp could not read back; and solve_all() takes the function saying whether this process writes the Solution and the State, which tssb_solver passes since it can run under MPI, so that the tool compiles again

• the MPBCfg.txt of ucblock_solver, tssb_solver, sddp_solver and investmentblock_solver, which every default configuration with a BundleSolver names, asked for Gurobi, so that these tools stopped at once where only HiGHS is there, e.g., in the conda packages: the master is now solved by HiGHS, Gurobi being one uncommented line away as the other :MILPSolver are

• sddp_solver solves an SDDPBlock whose scenarios come from a ScenarioGenerator, such as examples/SDDPBlock-new.nc4, also when the stages are solved by a LagrangianDualSolver: it set scenario 0 at every stage before attaching the Solver, i.e., before the pool of scenarios the SDDPSolver prepares when it is attached, and stopped with "invalid scenario index 0"; that was only there for OSIMPSolver, which BundleSolver no longer has, and it is gone

• ucblock_solver takes LagrangianDualSolver from its plain makefile, as the other tools do, rather than from the one that assumes the library was installed

• the duals of the pollutant constraints are read by pollutant and by zone again, a single index having mixed the zones of one pollutant with those of another in the output of the SDDP tool

• print_status() closes its parenthesis and goes to a new line whatever the status is, kLowPrecision and the ones below it having left the line open and run into what came next

• ucblock_solver and svm_solver set the log of their Solver as the other tools do, so that intLogVerb of a configuration is heard instead of being read and dropped

• the header of each file of the pollutant duals written by smspp_sddp_solver names the zones rather than repeating Zone_0

0.7.1 - 2026-09-14

Fixed

• the link that carries the name a tool had before the prefix is made in the directory of the install and not in the one of the configure, so that cmake --install --prefix puts it next to the tool instead of failing on the directory of the machine

0.7.0 - 2026-09-13

Changed

• the installed executables carry the name of the project, e.g. smspp_ucblock_solver and smspp_chgcfg, so that they are recognisable among all the others where they are installed; each of them is also installed under the name it had before, which is a link to it and which a later release will drop, and both names have their shell completions and their man page

0.6.0 - 2026-09-12

Added

• the configurations that attach the dynamic programming Solver of the nuclear units inside ucblock_solver and inside the timing study (NUBSCfg-DP.txt), the Lagrangian chain of a unit commitment whose units are nuclear being otherwise that of the thermal ones

• svm_solver, a SVMBlock solver that trains a Support Vector Machine and performs the model selection around it: hold-out and k-fold cross-validation, both stratified, and grid search over the hyper-parameters

• ml_utils.{h,cpp}, the model-agnostic machine learning scaffolding that svm_solver uses, i.e., the splits, the scores and the grid; it depends on nothing but the standard library, so any tool training a model can use it

• every tool installs its configuration, its example instances and a man page generated from its --help; without -c, when none of its configuration files is found with its own prefix, it uses the installed configuration

• the --help of every tool follows the GNU layout, and describes its input file, how the configuration files are looked up, some examples and the exit status

• an empty name given to -B or -S, as in -B '', means no file at all

• chgcfg has --help, --version and a man page, as every other tool

• mcfblock_solver, bkblock_solver, cflblock_solver, mmcfblock_solver and sfdcrblock_solver, the solvers of MCFBlock, BinaryKnapsackBlock, CapacitatedFacilityLocationBlock, MMCFBlock and SingleFlowDCRBlock, which read their Block from an SMS++ netCDF file or from a native text format of it, chosen by -f

Changed

• the default configurations of ucblock_solver, tssb_solver and sddp_solver solve with HiGHSMILPSolver, which needs no license

• the default -B of ucblock_solver is InnerBCfg.txt, which chooses the formulation of the units and of the network

• -n writes the problem on the file it is given, and -v takes its level only when attached, as in -v2 or --verbose=2, so that -v can precede the input file

• svm_solver takes the kernel from the command line with -K, one of linear, poly, gaussian, laplacian and sigmoid, which overrides the one the instance carries while keeping its parameters

• svm_solver reports the status and the bounds of the training problem in the format every other tool uses, and -O writes the trained model, i.e., the SVMBlockSolution

• svm_solver trains the models of a model selection in parallel, the grid and the folds being a cartesian product of independent problems; -j says how many at a time

• the splits of ml_utils are specified down to the bit, rather than being left to the implementation-defined shuffle of the standard library, so that a seed gives the same splits everywhere and an experiment can be reproduced in any other language

• the PPH configuration of ucblock_solver solves the Lagrangian Dual of every proximal iteration to convergence, and follows the parameters of PrimalProximalHeur being now named after the algorithm they belong to

• the version of the module is the git tag of its repository, or the VERSION.txt of a release tarball, and the shared library carries it: its SONAME is major.minor while the major is 0, and it is installed with an RPATH relative to itself, so that an installed tree keeps working wherever it is moved

Fixed

• svm_solver reads a netCDF SVMBlock again, and looks for the input file at the -p prefix whatever its format

• the tools report a missing input file and an unknown option on the standard error

• the shell completions of the tools name their tool, instead of nothing

0.5.4 - 2025-12-12

Added

• support fir set_solver_log() in all solvers

• std::set_terminate() support in all solvers

• MPI support to the computation of InvestmentFunction

• support for reading initial State and writing the final one

• support for dry runs

• support for reading initial Solution and writing the final one

Changed

• all things that can be changed, and the common definitions, are now in makefile_common to reduce code duplication within makefiles and to make adapting to one's environment quicker

• all *_solver executables share the same baisc set of command-line options, then can add upon it

• updated block solver handling in CutProcessing

Fixed

• consider UnitBlock scaling when outputting the solution

• get_installed_quantity in investment_solver

0.5.3 - 2024-02-29

Changed

• adapted to new CMake / makefile organisation

Removed

• the -r option from sddp_solver

0.5.2 - 2023-05-17

Added

• investment_solver keeps track of the best solution found.

• If the State file is not found, investment_solver shows a warning and proceeds.

• Save the best solution and the two most recent SolverState in investment_solver.

• Implement linear constraints in InvestmentFunction.

• InvestmentFunctionState.

Changed

• Disable the computation of linearizations in InvestmentFunction in simulation mode.

Fixed

• The initial up and downtimes of thermal units are only updated in investment_solver when a single scenario is being simulated.

• Linearization of InvestmentFunction.

0.5.1 - 2022-07-01

Added

• The investment_solver tool to solve an InvestmentBlock.

• The chgcfg tool to change configuration files.

• Consecutive simulations to sddp_solver.

0.5.0 - 2021-12-08

Added

• Multiple parameters to sddp_solver.

• MPI support to sddp_solver.

• Configuration of LagrangianDualSolver in sddp_solver.

Fixed

• Output of UCBlock solution.

• Initial conditions for simulation in sddp_solver.

0.4.0 - 2021-02-05

Added

• sddp_solver tool.

Changed

• Block/ucblock/thermalunit solvers have now the same interface.

• Major review of project tree.

0.3.1 - 2020-09-28

Fixed

• A bug in ucblock_solver that prevented configuration loading.

0.3.0 - 2020-09-16

Added

• Support for new configuration framework.

0.2.0 - 2020-03-06

Added

• Changelog.

Fixed

• Minor fixes.

0.1.0 - 2020-01-06

Added

• First test release.