commonpower.control.safety_layer.safety_layers.ActionProjectionSafetyLayer

class ActionProjectionSafetyLayer(penalty: BasePenalty, solver: OptSolver | None = None)[source]

Computes safe action by minimizing the distance between the RL action and the safe action while also satisfying constraints.

Parameters:

penalty (BasePenalty) – class defining the penalty behavior for unsafe actions
solver (OptSolver, optional) – solver for optimization problem, defaults to direct gurobi https://pyomo.readthedocs.io/en/stable/library_reference/solvers/gurobi_direct.html

Returns:

ActionReplacementSafetyLayer

Methods

`action_distance`	Euclidean norm between action expressed by the model and the action in dict
`compute_safe_action`	Checks whether the actions proposed by the controller satisfy the constraints of the controlled entities and replaces by pyomo-generated output if necessary.
`get_penalty`	Get penalty depending on the penalty class used.
`initialize`	Initializes the safety layer :param nodes: list of controlled entities to be safeguarded :type nodes: List[ModelEntity] :param top_level_nodes: list of controlled entities in highest level of model tree :type top_level_nodes: List[ModelEntity] :param solver: solver for optimization problem which will be called by Pyomo :type solver: OptSolver
`is_action_feasible`	As feasibility check is quite costly, we can skip for projection layer as we check action distance to the original action after optimization
`prepare_model`	Prepare model.
`set_action_from_model`	Corrects an action according to a model.
`set_action_in_model`	Sets the model values to the action to initialize the optimization.
`solve_model`	Finds a feasible action by solving a local model.

Attributes

DISTANCE_EPS

is_action_feasible(action) → bool[source]: As feasibility check is quite costly, we can skip for projection layer as we check action distance to the original action after optimization

prepare_model() → ConcreteModel[source]

Prepare model. Additionally projection criterion is added

Returns:

ConcreteModel –

pyomo optimization model, representing: the part of the system under supervision of the safety