gym_anm.agents.mpc.MPCAgent

class gym_anm.agents.mpc.MPCAgent(simulator, action_space, gamma, safety_margin=0.9, planning_steps=1)[source]

Bases: object

A base class for an agent that solves a MPC DC Optimal Power Flow.

This agent accesses the full state of the distribution network at time \(t\) and then solves an N-stage optimization problem. The optimization problem solved is an instance of a Direct Current (DC) Optimal Power Flow problem.

The construction of the DC OPF problem relies on 3 assumptions:

  1. transmission lines are lossless, i.e., \(r_{ij} = 0, \forall e_{ij} \in \mathcal E,\)

  2. the difference between adjacent bus voltage angles is small, i.e., \(\angle V_i = \angle V_j, \forall e_{ij} \in \mathcal E\),

  3. nodal voltage magnitude are close to unity, i.e., \(|V_{i}| = 1, \forall i \in \mathcal N\).

The construction of the N-stage optimization problem relies on two additional assumptions:

All sub-classes must implement the :py:func`forecast()` method to provide forecasts of demand and generation over the optimization horizon \([t+1,t+N]\).

All values are used in per-unit.

__init__(simulator, action_space, gamma, safety_margin=0.9, planning_steps=1)[source]
Parameters:

  • simulator (gym_anm.simulator.simulator.Simulator) – The electricity distribution network simulator.

  • action_space (gym.spaces.Box) – The action space of the environment (used to clip actions).

  • gamma (float) – The discount factor in [0, 1].

  • safety_margin (float, optional) – The safety margin constant \(\beta\) in [0, 1], used to further constraint the power flow on each transmission line, thus likely accounting for the error introduced in the DC approximation.

  • planning_steps (int, optional) – The number (N) of stages (time steps) taken into account in the optimization problem.

Methods

__init__(simulator, action_space, gamma[, ...])

param simulator:

The electricity distribution network simulator.

act(env)

Select an action by solving the N-stage DC OPF.

forecast(env)

Forecast the demand and generation over the optimization horizon.
act(env)[source]

Select an action by solving the N-stage DC OPF.

Parameters:

env (py:class:gym_anm.ANMEnv) – The gym-anm environment.

Returns:

The action vector to apply in the environment.

Return type:

numpy.ndarray

forecast(env)[source]

Forecast the demand and generation over the optimization horizon.

This method must be implemented by all sub-classes of MPCAgent. It gets called in act() and must return the predictions of future load demand, \(\tilde P_{l,k}^{(dev)}\), and maximum generation at non-slack generators, \(\tilde P_{g,k}^{(max)}\).

Parameters:

env (py:class:gym_anm.ANMEnv) – The gym-anm environment.

Returns:

  • P_load_forecast (array_like) – A (N_load, N) array of forecasted load power injections (<0), where N is the length of the optimization horizon. The rows should be ordered in increasing order of device ID.

  • P_gen_forecast (array_like) – A (N_gen-1, N) array of forecasted maximum generation from non-slack generators, where N is the length of the optimization horizon. The rows should be ordered in increasing order of device ID.