See the module pypsa.opf.
Optimisation with the full non-linear power flow equations is not yet supported.
Optimisation with the linearised power flow equations for (mixed) AC and DC networks is fully supported.
All constraints and variables are listed below.
Execute:
network.lopf(snapshots, solver_name="glpk", solver_io=None, extra_functionality=None, solver_options={}, keep_files=False, formulation="angles")
where snapshots is an iterable of snapshots, solver_name is a
string, e.g. "gurobi" or "glpk", solver_io is a string,
extra_functionality is a function of network and snapshots that is
called before the solver (see below), solver_options is a
dictionary of flags to pass to the solver, keep_files means that
the .lp file is saved and formulation is a string in
["angles","cycles","kirchhoff","ptdf"] (see :ref:`formulations` for more details).
The linear OPF module can optimises the dispatch of generation and storage and the capacities of generation, storage and transmission.
It is assumed that the load is inelastic and must be met in every snapshot (this will be relaxed in future versions).
The optimisation currently uses continuous variables for most functionality; unit commitment with binary variables is also implemented for generators.
The objective function is the total system cost for the snapshots optimised.
Each snapshot can be given a weighting w_t to represent e.g. multiple hours.
This set-up can also be used for stochastic optimisation, if you interpret the weighting as a probability.
Each transmission asset has a capital cost.
Each generation and storage asset has a capital cost and a marginal cost.
WARNING: If the transmission capacity is changed in passive networks, then the impedance will also change (i.e. if parallel lines are installed). This is NOT reflected in the LOPF, so the network equations may no longer be valid. Note also that all the expansion is continuous.
Capacity optimisation can be turned off so that only the dispatch is optimised, like a short-run electricity market model.
For simplified transmission representation using Net Transfer Capacities (NTCs), there is a Link component which does controllable power flow like a transport model (and can also represent a point-to-point HVDC link).
To minimise long-run annual system costs for meeting an inelastic electrical load, capital costs for transmission and generation should be set to the annualised investment costs in e.g. EUR/MW/a, marginal costs for dispatch to e.g. EUR/MWh and the weightings (now with units hours per annum, h/a) are chosen such that
\sum_t w_t = 8760
In this case the objective function gives total system cost in EUR/a to meet the total load.
For the very simplest stochastic optimisation you can use the
weightings w_t as probabilities for the snapshots, which can
represent different load/weather conditions. More sophisticated
functionality is planned.
n \in N = \{0,\dots |N|-1\} label the buses
t \in T = \{0,\dots |T|-1\} label the snapshots
l \in L = \{0,\dots |L|-1\} label the branches
s \in S = \{0,\dots |S|-1\} label the different generator/storage types at each bus
w_t weighting of time t in the objective function
g_{n,s,t} dispatch of generator s at bus n at time t
\bar{g}_{n,s} nominal power of generator s at bus n
\bar{g}_{n,s,t} availability of generator s at bus n at time t per unit of nominal power
u_{n,s,t} binary status variable for generator with unit commitment
suc_{n,s,t} start-up cost if generator with unit commitment is started at time t
sdc_{n,s,t} shut-down cost if generator with unit commitment is shut down at time t
c_{n,s} capital cost of extending generator nominal power by one MW
o_{n,s} marginal cost of dispatch generator for one MWh
f_{l,t} flow of power in branch l at time t
F_{l} capacity of branch l
\eta_{n,s} efficiency of generator s at bus n
\eta_{l} efficiency of controllable link l
e_s CO2-equivalent-tonne-per-MWh of the fuel carrier s
Further definitions are given below.
See pypsa.opf.define_linear_objective(network,snapshots).
The objective function is composed of capital costs c for each component and operation costs o for generators
\sum_{n,s} c_{n,s} \bar{g}_{n,s} + \sum_{n,s} c_{n,s} \bar{h}_{n,s} + \sum_{l} c_{l} F_l \\
+ \sum_{t} w_t \left[\sum_{n,s} o_{n,s,t} g_{n,s,t} + \sum_{n,s} o_{n,s,t} h_{n,s,t} \right]
+ \sum_{t} \left[suc_{n,s,t} + sdc_{n,s,t} \right]
Additional variables which do not appear in the objective function are the storage uptake variable, the state of charge and the voltage angle for each bus.
These are defined in pypsa.opf.define_generator_variables_constraints(network,snapshots).
Generator nominal power and generator dispatch for each snapshot may be optimised.
Each generator has a dispatch variable g_{n,s,t} where n labels the bus, s labels the particular generator at the bus (e.g. it can represent wind/gas/coal generators at the same bus in an aggregated network) and t labels the time.
It obeys the constraints:
\tilde{g}_{n,s,t}*\bar{g}_{n,s} \leq g_{n,s,t} \leq \bar{g}_{n,s,t}*\bar{g}_{n,s}
where \bar{g}_{n,s} is the nominal power (generator.p_nom)
and \tilde{g}_{n,s,t} and \bar{g}_{n,s,t} are
time-dependent restrictions on the dispatch (per unit of nominal
power) due to e.g. wind availability or power plant de-rating.
For generators with time-varying p_max_pu in network.generators_t the per unit
availability \bar{g}_{n,s,t} is a time series.
For generators with static p_max_pu in network.generators the per unit
availability is a constant.
If the generator's nominal power \bar{g}_{n,s} is also the
subject of optimisation (generator.p_nom_extendable == True) then
limits generator.p_nom_min and generator.p_nom_max on the
installable nominal power may also be introduced, e.g.
\tilde{g}_{n,s} \leq \bar{g}_{n,s} \leq \hat{g}_{n,s}
These are defined in pypsa.opf.define_generator_variables_constraints(network,snapshots).
The implementation follows Chapter 4.3 of Convex Optimization of Power Systems by Joshua Adam Taylor (CUP, 2015).
Unit commitment can be turned on for any generator by setting committable to be True. This introduces a
times series of new binary status variables u_{n,s,t} \in \{0,1\},
which indicates whether the generator is running (1) or not (0) in
period t. The restrictions on generator output now become:
u_{n,s,t}*\tilde{g}_{n,s,t}*\bar{g}_{n,s} \leq g_{n,s,t} \leq u_{n,s,t}*\bar{g}_{n,s,t}*\bar{g}_{n,s} \hspace{.5cm} \forall\, n,s,t
so that if u_{n,s,t} = 0 then also g_{n,s,t} = 0.
If T_{\textrm{min\_up}} is the minimum up time then we have
\sum_{t'=t}^{t+T_\textrm{min\_up}} u_{n,s,t'}\geq T_\textrm{min\_up} (u_{n,s,t} - u_{n,s,t-1}) \hspace{.5cm} \forall\, n,s,t
(i.e. if the generator has just started up (u_{n,s,t} - u_{n,s,t-1} = 1) then it has to run for at least T_{\textrm{min\_up}} periods). Similarly for a minimum down time of T_{\textrm{min\_down}}
\sum_{t'=t}^{t+T_\textrm{min\_down}} (1-u_{n,s,t'})\geq T_\textrm{min\_down} (u_{n,s,t-1} - u_{n,s,t}) \hspace{.5cm} \forall\, n,s,t
For non-zero start up costs suc_{n,s} a new variable suc_{n,s,t} \geq 0 is introduced for each time period t and added to the objective function. The variable satisfies
suc_{n,s,t} \geq suc_{n,s} (u_{n,s,t} - u_{n,s,t-1}) \hspace{.5cm} \forall\, n,s,t
so that it is only non-zero if u_{n,s,t} - u_{n,s,t-1} = 1, i.e. the generator has just started, in which case the inequality is saturated suc_{n,s,t} = suc_{n,s}. Similarly for the shut down costs sdc_{n,s,t} \geq 0 we have
sdc_{n,s,t} \geq sdc_{n,s} (u_{n,s,t-1} - u_{n,s,t}) \hspace{.5cm} \forall\, n,s,t
These are defined in pypsa.opf.define_generator_variables_constraints(network,snapshots).
The implementation follows Chapter 4.3 of Convex Optimization of Power Systems by Joshua Adam Taylor (CUP, 2015).
Ramp rate limits can be defined for increasing power output ru_{n,s} and decreasing power output rd_{n,s}. By default these are null and ignored. They should be given per unit of the generator nominal power. The generator dispatch then obeys
-rd_{n,s} * \bar{g}_{n,s} \leq (g_{n,s,t} - g_{n,s,t-1}) \leq ru_{n,s} * \bar{g}_{n,s}
for t \in \{1,\dots |T|-1\}.
For generators with unit commitment you can also specify ramp limits at start-up rusu_{n,s} and shut-down rdsd_{n,s}
\left[ -rd_{n,s}*u_{n,s,t} -rdsd_{n,s}(u_{n,s,t-1} - u_{n,s,t})\right] \bar{g}_{n,s} \leq (g_{n,s,t} - g_{n,s,t-1}) \leq \left[ru_{n,s}*u_{n,s,t-1} + rusu_{n,s} (u_{n,s,t} - u_{n,s,t-1})\right]\bar{g}_{n,s}
These are defined in pypsa.opf.define_storage_variables_constraints(network,snapshots).
Storage nominal power and dispatch for each snapshot may be optimised.
With a storage unit the maximum state of charge may not be independently optimised from the maximum power output (they're linked by the maximum hours variable) and the maximum power output is linked to the maximum power input. To optimise these capacities independently, build a storage unit out of the more fundamental Store and Link components.
The storage nominal power is given by \bar{h}_{n,s}.
In contrast to the generator, which has one time-dependent variable, each storage unit has three:
The storage dispatch h_{n,s,t} (when it depletes the state of charge):
0 \leq h_{n,s,t} \leq \bar{h}_{n,s}
The storage uptake f_{n,s,t} (when it increases the state of charge):
0 \leq f_{n,s,t} \leq \bar{h}_{n,s}
and the state of charge itself:
0\leq soc_{n,s,t} \leq r_{n,s} \bar{h}_{n,s}
where r_{n,s} is the number of hours at nominal power that fill the state of charge.
The variables are related by
soc_{n,s,t} = \eta_{\textrm{stand};n,s}^{w_t} soc_{n,s,t-1} + \eta_{\textrm{store};n,s} w_t f_{n,s,t} - \eta^{-1}_{\textrm{dispatch};n,s} w_t h_{n,s,t} + w_t\textrm{inflow}_{n,s,t} - w_t\textrm{spillage}_{n,s,t}
\eta_{\textrm{stand};n,s} is the standing losses dues to e.g. thermal losses for thermal storage. \eta_{\textrm{store};n,s} and \eta_{\textrm{dispatch};n,s} are the efficiency losses for power going into and out of the storage unit.
There are two options for specifying the initial state of charge soc_{n,s,t=-1}: you can set
storage_unit.cyclic_state_of_charge = False (the default) and the value of
storage_unit.state_of_charge_initial in MWh; or you can set
storage_unit.cyclic_state_of_charge = True and then
the optimisation assumes soc_{n,s,t=-1} = soc_{n,s,t=|T|-1}.
If in the time series storage_unit_t.state_of_charge_set there are
values which are not NaNs, then it will be assumed that these are
fixed state of charges desired for that time t and these will
be added as extra constraints. (A possible usage case would be a
storage unit where the state of charge must empty every day.)
These are defined in pypsa.opf.define_store_variables_constraints(network,snapshots).
Store nominal energy and dispatch for each snapshot may be optimised.
The store nominal energy is given by \bar{e}_{n,s}.
The store has two time-dependent variables:
The store dispatch h_{n,s,t}:
-\infty \leq h_{n,s,t} \leq +\infty
and the energy:
\tilde{e}_{n,s} \leq e_{n,s,t} \leq \bar{e}_{n,s}
The variables are related by
e_{n,s,t} = \eta_{\textrm{stand};n,s}^{w_t} e_{n,s,t-1} - w_t h_{n,s,t}
\eta_{\textrm{stand};n,s} is the standing losses dues to e.g. thermal losses for thermal storage.
There are two options for specifying the initial energy
e_{n,s,t=-1}: you can set
store.e_cyclic = False (the default) and the
value of store.e_initial in MWh; or you can
set store.e_cyclic = True and then the
optimisation assumes e_{n,s,t=-1} = e_{n,s,t=|T|-1}.
See pypsa.opf.define_passive_branch_flows(network,snapshots) and
pypsa.opf.define_passive_branch_constraints(network,snapshots) and pypsa.opf.define_branch_extension_variables(network,snapshots).
For lines and transformers, whose power flows according to impedances, the power flow f_{l,t} in AC networks is given by the difference in voltage angles \theta_{n,t} at bus0 and \theta_{m,t} at bus1 divided by the series reactance x_l
f_{l,t} = \frac{\theta_{n,t} - \theta_{m,t}}{x_l}
(For DC networks, replace the voltage angles by the difference in voltage magnitude \delta V_{n,t} and the series reactance by the series resistance r_l.)
This flow is the limited by the capacity :math:F_l of the line
|f_{l,t}| \leq F_l
Note that if F_l is also subject to optimisation
(branch.s_nom_extendable == True), then the impedance x of
the line is NOT automatically changed with the capacity (to represent
e.g. parallel lines being added).
There are two choices here:
Iterate the LOPF again with the updated impedances (see e.g. http://www.sciencedirect.com/science/article/pii/S0360544214000322#).
João Gorenstein Dedecca has also implemented a MILP version of the transmission expansion, see https://github.com/jdedecca/MILP_PyPSA, which properly takes account of the impedance with a disjunctive relaxation. This will be pulled into the main PyPSA code base soon.
PyPSA implements four formulations of the linear power flow equations that are mathematically equivalent, but may have different solving times. These different formulations are described and benchmarked in the arXiv preprint paper Linear Optimal Power Flow Using Cycle Flows.
You can choose the formulation by passing network.lopf the
argument formulation, which must be in
["angles","cycles","kirchhoff","ptdf"]. angles is the standard
formulations based on voltage angles described above, used for the
linear power flow and found in textbooks. ptdf uses the Power
Transfer Distribution Factor (PTDF) formulation, found for example in
http://www.sciencedirect.com/science/article/pii/S0360544214000322#. kirchhoff
and cycles are two new formulations based on a graph-theoretic
decomposition of the network flows into a spanning tree and closed
cycles.
Based on the benchmarking in Linear Optimal Power Flow Using Cycle
Flows for standard networks,
kirchhoff almost always solves fastest, averaging 3 times faster
than the angles formulation and up to 20 times faster in specific
cases. The speedup is higher for larger networks with dispatchable
generators at most nodes.
See pypsa.opf.define_controllable_branch_flows(network,snapshots)
and pypsa.opf.define_branch_extension_variables(network,snapshots).
For links, whose power flow is controllable, there is simply an optimisation variable for each component which satisfies
|f_{l,t}| \leq F_l
If the link flow is positive f_{l,t} > 0 then it withdraws
f_{l,t} from bus0 and feeds in \eta_l f_{l,t} to
bus1, where \eta_l is the link efficiency.
If additional output buses busi for i=2,3,\dots are
defined (i.e. bus2, bus3, etc) and their associated
efficiencies efficiencyi, i.e. \eta_{i,l}, then at
busi the feed-in is \eta_{i,l} f_{l,t}. See also
:ref:`components-links-multiple-outputs`.
See pypsa.opf.define_nodal_balances(network,snapshots).
This is the most important equation, which guarantees that the power balances at each bus n for each time t.
\sum_{s} g_{n,s,t} + \sum_{s} h_{n,s,t} - \sum_{s} f_{n,s,t} - \sum_{l} K_{nl} f_{l,t} = \sum_{s} d_{n,s,t} \hspace{.4cm} \leftrightarrow \hspace{.4cm} w_t\lambda_{n,t}
Where d_{n,s,t} is the exogenous load at each node (load.p_set) and the incidence matrix K_{nl} for the graph takes values in \{-1,0,1\} depending on whether the branch l ends or starts at the bus. \lambda_{n,t} is the shadow price of the constraint, i.e. the locational marginal price, stored in network.buses_t.marginal_price.
The bus's role is to enforce energy conservation for all elements feeding in and out of it (i.e. like Kirchhoff's Current Law).
See pypsa.opf.define_global_constraints(network,snapshots).
Global constraints apply to more than one component.
Currently only "primary energy" constraints are defined. They depend on the power plant efficiency and carrier-specific attributes such as specific CO2 emissions.
Suppose there is a global constraint defined for CO2 emissions with
sense <= and constant \textrm{CAP}_{CO2}. Emissions can come
from generators whose energy carriers have CO2 emissions and from
stores and storage units whose storage medium releases or absorbs CO2
when it is converted. Only stores and storage units with non-cyclic
state of charge that is different at the start and end of the
simulation can contribute.
If the specific emissions of energy carrier s is e_s
(carrier.co2_emissions) CO2-equivalent-tonne-per-MWh and the
generator with carrier s at node n has efficiency
\eta_{n,s} then the CO2 constraint is
\sum_{n,s,t} \frac{1}{\eta_{n,s}} w_t\cdot g_{n,s,t}\cdot e_{n,s} + \sum_{n,s}\left(e_{n,s,t=-1} - e_{n,s,t=|T|-1}\right) \cdot e_{n,s} \leq \textrm{CAP}_{CO2} \hspace{.4cm} \leftrightarrow \hspace{.4cm} \mu
The first sum is over generators; the second sum is over stores and
storage units. \mu is the shadow price of the constraint,
i.e. the CO2 price in this case. \mu is an output of the
optimisation stored in network.global_constraints.mu.
PyPSA uses the Python optimisation language pyomo to construct the OPF problem. You can easily
extend the optimisation problem constructed by PyPSA using the usual
pyomo syntax. To do this, pass the function network.lopf a
function extra_functionality as an argument. This function must
take two arguments extra_functionality(network,snapshots) and is
called after the model building is complete, but before it is sent to
the solver. It allows the user to add, change or remove constraints
and alter the objective function.
The CHP example and the
example that replaces generators and storage units with fundamental links
and stores
both pass an extra_functionality argument to the LOPF to add
functionality.
For the linear optimal power flow, the following data for each component are used. For almost all values, defaults are assumed if not explicitly set. For the defaults and units, see :doc:`components`.
network{snapshot_weightings}
bus.{v_nom, carrier}
load.{p_set}
generator.{p_nom, p_nom_extendable, p_nom_min, p_nom_max, p_min_pu, p_max_pu, marginal_cost, capital_cost, efficiency, carrier}
storage_unit.{p_nom, p_nom_extendable, p_nom_min, p_nom_max, p_min_pu, p_max_pu, marginal_cost, capital_cost, efficiency*, standing_loss, inflow, state_of_charge_set, max_hours, state_of_charge_initial, cyclic_state_of_charge}
store.{e_nom, e_nom_extendable, e_nom_min, e_nom_max, e_min_pu, e_max_pu, e_cyclic, e_initial, capital_cost, marginal_cost, standing_loss}
line.{x, s_nom, s_nom_extendable, s_nom_min, s_nom_max, capital_cost}
transformer.{x, s_nom, s_nom_extendable, s_nom_min, s_nom_max, capital_cost}
link.{p_min_pu, p_max_pu, p_nom, p_nom_extendable, p_nom_min, p_nom_max, capital_cost}
carrier.{carrier_attribute}
global_constraint.{type, carrier_attribute, sense, constant}
Note that for lines and transformers you MUST make sure that x is non-zero, otherwise the bus admittance matrix will be singular.
bus.{v_mag_pu, v_ang, p, marginal_price}
load.{p}
generator.{p, p_nom_opt}
storage_unit.{p, p_nom_opt, state_of_charge, spill}
store.{p, e_nom_opt, e}
line.{p0, p1, s_nom_opt, mu_lower, mu_upper}
transformer.{p0, p1, s_nom_opt, mu_lower, mu_upper}
link.{p0, p1, p_nom_opt, mu_lower, mu_upper}
global_constraint.{mu}