Support vector machine enhanced models for short-term wind farm generation forecasting
Systems and methods for forecasting wind farm power generation are disclosed. Via use of a support vector machine (SVM) enhanced Markov model, short-term wind power generation forecasts may be generated. Exemplary approaches accurately account for wind ramp-up and ramp-down, as well as diurnal non-stationarity and seasonality of wind power generation. Via use of the disclosed forecasting approaches, utilities and grid managers can make improved decisions relating to electrical power generation and transmission, thus reducing costs and reducing pollution.
1. A method for improving the operation of a fossil fuel power plant, the method comprising:
deploying a meteorological tower at a location co-located with a wind turbine in a wind farm, wherein the wind farm comprises a plurality of wind turbines, wherein the wind farm is coupled to an electrical grid, and wherein the fossil fuel power plant is coupled to the electrical grid;
transmitting from the meteorological tower and to a processor for forecasting wind farm power, wind speed information associated with the wind turbine;
receiving, at the processor and from a wind farm power generation sensor coupled to the wind farm, power generation information for the wind farm;
identifying, by the processor, relationships among the wind turbines in the wind farm using minimum spanning trees;
calculating, by the processor and using the minimum spanning trees, power output relationships among the wind turbines;
creating, by the processor, a finite state space Markov chain forecast model for the wind turbines in the wind farm;
creating, by the processor, a support vector machine (SVM) model for each state in the Markov chain;
integrating, by the processor, the SVM model into the Markov model to generate a forecast of the wind farm power generation;
transmitting, by the processor, the forecast of the wind farm power generation to a manager of the electrical grid; and
based on the forecast of the wind farm power generation, reducing, by the manager of the electrical grid, excessive fossil fuel consumption arising from undesired excess electrical generation at the fossil fuel power plant.
2. The method of claim 1 , wherein the forecast of the wind farm power generation is generated using the equation
{circumflex over (P)} ag SVM ( t +1)= P ag ( t )+ R j* k ,
wherein P ag (t)∈[Γ k , Γ k+1 ) is the current observed wind farm power generation, wherein the corresponding forecast state S {circumflex over (k)} is the state satisfying {circumflex over (P)} ag SVM (t+1)∈[Γ {circumflex over (k)} ,Γ {circumflex over (k)}+1 ), and wherein:
S {circumflex over (k)} is a forecast state using the SVM model;
Γ k is a wind farm generation level;
P ag (t) is the aggregate power output of the wind farm at time t; and
{circumflex over (P)} ag SVM is the forecast of the wind farm power generation.
3. The method of claim 2 , wherein the forecast of the wind farm power generation is a distributional forecast, and wherein the distributional forecast is generated using the equation
Pr
(
P
ag
(
t
+
1
)
=
P
ag
,
j
❘
S
(
t
)
=
S
k
,
x
(
t
)
)
=
{
q
k
;
if
j
=
k
^
;
(
1
-
q
k
)
Q
k
j
∑
l
≠
k
^
Q
kl
,
otherwise
,
wherein:
S(t) is the state of the Markov chain at time t;
x(t)∈ l is the observed wind farm power generation at time t;
q k is the transition probability from S k to S {circumflex over (k)} ; and
Q is a transition matrix of the Markov chain.
4. The method of claim 3 , further comprising generating, by the processor, a point forecast from the distributional forecast, and wherein the point forecast is generated using the equation
P
^
ag
(
t
+
1
)
=
P
ag
,
k
^
q
k
+
∑
j
≠
k
^
P
ag
,
j
(
1
-
q
k
)
Q
kj
∑
l
≠
k
^
Q
kl
.
5. The method of claim 1 , wherein each minimum spanning tree:
comprises a subset of the wind turbines in the wind farm;
identifies a parent turbine for each wind turbine in the subset; and
guarantees that each wind turbine in the subset is linked to a root turbine of the subset.
6. The method of claim 5 , wherein the wind turbine is the root turbine in a minimum spanning tree.
7. The method of claim 1 , wherein the creating, b the processor, a support vector machine (SVM) model for each state in the Markov chain is performed in parallel to improve computational efficiency.
8. The method of claim 1 , wherein the forecast of the wind farm power generation indicates the presence of a wind ramp.
9. The method of claim 1 , wherein the creation of the finite state space Markov chain forecast model utilizes historical data comprising at least one of historical data for wind farm power generation or historical data for wind speed from the meteorological tower.
10. The method of claim 1 , wherein the creation of the finite state space Markov chain forecast model utilizes ramp trend information.
11. The method of claim 1 , wherein the forecast of the wind farm power generation predicts the power output of the wind farm for a period of between 5 minutes and 1 hour into the future.