Chance constrained extreme learning machine method for nonparametric interval forecasting of wind power
The present application discloses a chance constrained extreme learning machine method for nonparametric interval forecasting of wind power, which belongs to the field of renewable energy generation forecasting. The method combines an extreme learning machine with a chance constrained optimization model, ensures that the interval coverage probability is no less than the confidence level by chance constraint, and takes minimizing the interval width as the training objective. The method avoids relying on the probability distribution hypothesis or limiting the interval boundary quantile level, so as to directly construct prediction intervals with well reliability and sharpness. The present application also proposes a bisection search algorithm based on difference of convex functions optimization to achieve efficient training for the chance constrained extreme learning machine.
1. A chance constrained extreme learning machine method for nonparametric interval forecasting of wind power, comprising the following steps of:
1) constructing a model of a chance constrained extreme learning machine;
2) constructing a sample average approximation model of the chance constrained extreme learning machine;
3) constructing a parametric 0-1 loss minimization model;
4) constructing a parametric difference of convex functions optimization model;
5) adopting a bisection search algorithm based on difference of convex functions optimization to train the chance constrained extreme learning machine;
wherein the step 1) comprises: comprehensively considering a joint probability distribution of a wind power and an input feature thereof, limiting the wind power to fall into prediction intervals with a probability not lower than a nominal confidence level by using chance constraint, and taking an expectation of minimizing an interval width as a training objective, and constructing the chance constrained extreme learning machine model:
min
ω
ℓ
,
ω
u
𝔼
μ
[
f
(
x
,
ω
u
)
-
f
(
x
,
ω
ℓ
)
]
which is subject to:
ℙ
μ
[
f
(
x
,
ω
ℓ
)
≤
y
≤
f
(
x
,
ω
u
)
]
≥
100
(
1
-
β
)
%
0
≤
f
(
x
,
ω
ℓ
)
≤
f
(
x
,
ω
u
)
≤
1
where x is a random variable corresponding to the input feature, y is a random variable corresponding to normalized wind power, the joint probability distribution of the two is denoted as μ(x,y); f(x, ) and f(x,ω u ) are output equations of the chance constrained extreme learning machine, which represent upper and lower boundaries of the prediction intervals, respectively, and ω u are weight vectors from a hidden layer of the extreme learning machine to an output neuron; 100 (1−β) % is the nominal confidence level of the prediction intervals; and denote expectation and probability operators, respectively;
the step 2) comprises: replacing the joint probability distribution of the input feature and wind power with an empirical probability distribution of training set samples thereof, approximating an actual expectation in an objective function by empirical expectation, and approximating an actual probability in the chance constraint by empirical probability to obtain the sample average approximate model of the chance constrained extreme learning machine;
v
★
=
min
γ
t
,
ω
ℓ
,
ω
u
∑
t
∈
𝒯
[
f
(
x
t
,
ω
u
)
-
f
(
x
t
,
ω
ℓ
)
]
which is constrained by:
γ
t
=
max
{
f
(
x
t
,
ω
ℓ
)
-
y
t
,
y
t
-
f
(
x
t
,
ω
u
)
}
,
∀
t
∈
𝒯
∑
t
∈
𝒯
[
1
-
𝕀
(
γ
t
≤
0
)
]
≤
β
❘
"\[LeftBracketingBar]"
𝒯
❘
"\[RightBracketingBar]"
0
≤
f
(
x
t
,
ω
ℓ
)
≤
f
(
x
t
,
ω
u
)
≤
1
,
∀
t
∈
𝒯
where v* is an optimal value of the optimization model, which denotes a shortest overall width of the prediction intervals satisfying the chance constraint; x t and y t are an input feature and the wind power; is a subscript set of various samples of the training set {(x t ,y t )} , | | is a number of the samples of the training set, β| | is a maximum number of wind power samples outside the prediction intervals at the nominal confidence level; Y t is an auxiliary variable indicating whether the wind power falls into the prediction intervals, with a non-negative value indicating that the wind power falls into a corresponding prediction interval of the prediction intervals or a positive value indicating that the wind power does not fall into a corresponding prediction interval of the prediction intervals; max {⋅} is a maximum function, taking a maximum value of each variable thereof; (⋅) is an indicator function of a logical discriminant, with a value being 1 when the logical discriminant is true and 0 when the logical discriminant is false;
the step 3) comprises: introducing a virtual parametric variable to represent an overall width budget of the prediction intervals, minimizing the probability that the wind power does not fall into the prediction intervals on a premise of meeting the overall width budget of the prediction intervals to obtain the parametric minimized 0-1 loss minimization model;
ρ
(
v
)
=
min
γ
t
,
ω
ℓ
,
ω
u
∑
t
∈
𝒯
[
1
-
𝕀
(
γ
t
≤
0
)
]
which is subject to:
γ
t
≥
f
(
x
t
,
ω
ℓ
)
-
y
t
,
∀
t
∈
𝒯
γ
t
≥
y
t
-
f
(
x
t
,
ω
u
)
,
∀
t
∈
𝒯
∑
t
∈
𝒯
[
f
(
x
t
,
ω
u
)
-
f
(
x
t
,
ω
ℓ
)
]
≤
v
0
≤
f
(
x
t
,
ω
ℓ
)
≤
f
(
x
t
,
ω
u
)
≤
1
,
∀
t
∈
𝒯
where v is the introduced virtual parametric variable representing the overall width budget of the prediction intervals; ρ(v) is an optimal value function of the minimized 0-1 loss model with respect to the parametric variable v, and a smallest parametric variable that satisfies the condition ρ(v)≤β| | is a shortest overall width v* of the prediction intervals that satisfies the chance constraint;
the step 4) comprises: approximating the indicator function of the logical discriminant in the objective function of the minimized 0-1 loss model with the parametric variable v by a difference of convex functions to obtain the difference of convex functions optimization model with the parametric variable v:
ρ
_
(
v
)
=
min
γ
t
,
ω
ℓ
,
ω
u
∑
t
∈
𝒯
[
1
-
L
DC
(
γ
t
;
m
)
]
which is constrained by:
L
DC
(
γ
t
;
m
)
=
max
{
-
m
γ
t
,
0
}
-
max
{
-
m
γ
t
-
1
,
0
}
,
∀
t
∈
𝒯
γ
t
≥
f
(
x
t
,
ω
ℓ
)
-
y
t
,
∀
t
∈
𝒯
γ
t
≥
y
t
-
f
(
x
t
,
ω
u
)
,
∀
t
∈
𝒯
∑
t
∈
𝒯
[
f
(
x
t
,
ω
u
)
-
f
(
x
t
,
ω
ℓ
)
]
≤
v
0
≤
f
(
x
t
,
ω
ℓ
)
≤
f
(
x
t
,
ω
u
)
≤
1
,
∀
t
∈
𝒯
where ρ (v) is an optimal value function of the difference of convex functions optimization model with the parameter v; L DC (γ t ; m) is a difference of convex functions approximating the indicator function (γ t ≤0); m is a positive slope parameter of the difference of convex functions, m is a positive value, and the greater the value of m, the higher the similarity between the difference of convex functions L DC (γ t ; m) and the indicator function (γ t ≤0); the decision vector of the model is denoted as θ=[ ] τ , where γ=[γ 1 γ 2 ⋅ ⋅ ⋅ ] τ ;
the step 5) comprises: using the bisection search algorithm based on difference of convex functions optimization to search for the shortest overall width v* of the prediction intervals that satisfy the chance constraint, so as to realize the training of the extreme learning machine; and further comprising the following steps:
step (1), giving a bisection search algorithm precision ∈ 1 and a bisection search interval [v l , v u ], wherein the given bisection search interval should contain the shortest overall width v* of the prediction intervals;
Step (2): for the parametric difference of convex functions optimization model, setting the parameter v thereof as a midpoint (v l +v u )/2 of the bisection search interval, and solving the difference of convex functions optimization model by using a convex-concave procedure algorithm, with a solution being θ =[ ] τ ;
step (3): calculating the number miss of samples of which the wind power in the training set falls outside the prediction intervals:
miss
←
∑
t
∈
𝒯
[
1
-
𝕀
(
f
(
x
t
,
ω
_
ℓ
)
≤
y
t
≤
f
(
x
t
,
ω
_
u
)
)
]
step (4), if miss ≤γ| |, updating the upper boundary v u ←(v l +v u )/2 of the bisection search interval and recording weight vectors ω l ← ω l , ω u ← ω u of an output layer of the current extreme learning machine; otherwise, updating the lower boundary v l ←(v l +v u )/2 of the bisection search interval;
step (5): if v u −v l ≤ε 1 , outputting the weight vectors ω l and ω u of the output layer of the extreme learning machine, otherwise, returning to step (2); and
6) constructing the prediction intervals {[f(x t ,ω l ),f (x t ,ω u )]} of a test set ={(x t ,y t )} based on the trained chance constrained extreme learning machine, evaluating the reliability of the prediction intervals using average coverage deviation (ACD), evaluating the sharpness of the prediction intervals using average width (AW), outputting the wind power with the minimized interval width, and providing more accurate uncertainty quantification information for decision makers of a power system.
2. The chance constrained extreme learning machine method for nonparametric interval forecasting of wind power according to claim 1 , wherein a convex-concave procedure algorithm is adopted to solve the difference of convex functions optimization model with a given parameter, which comprises the following steps:
step (1): giving an algorithm convergence accuracy ∈ 2 , the slope parameter m of the difference of convex functions, and the parameter v representing the overall width budget of the prediction intervals;
step (2): setting an iteration counter k←0; solving the following linear programming problem to obtain an initial solution θ (0) of the difference of convex functions optimization model:
θ
(
0
)
←
arg
min
γ
t
,
ω
ℓ
,
ω
u
1
T
γ
which is subject to:
γ
t
≥
f
(
x
t
,
ω
ℓ
)
-
y
t
,
∀
t
∈
𝒯
γ
t
≥
y
t
-
f
(
x
t
,
ω
u
)
,
∀
t
∈
𝒯
∑
t
∈
𝒯
[
f
(
x
t
,
ω
u
)
-
f
(
x
t
,
ω
ℓ
)
]
≤
v
0
≤
f
(
x
t
,
ω
ℓ
)
≤
f
(
x
t
,
ω
u
)
≤
1
,
∀
t
∈
𝒯
where 1 is a vector with all elements being 1, whose dimension is the same as the number of the samples in the training set;
step (3): updating the solution of the parametric difference of convex functions optimization model in a (k+1) th iteration by using the following formula:
θ
(
k
+
1
)
←
arg
min
γ
t
,
ω
ℓ
,
ω
u
L
vex
+
(
γ
)
-
[
L
vex
-
(
γ
(
k
)
)
+
δ
(
k
)
T
(
γ
-
γ
(
k
)
)
]
which
is
subject
to
:
L
vex
+
(
γ
)
=
∑
t
∈
𝒯
[
1
+
max
{
-
m
γ
t
-
1
,
0
}
]
L
vex
-
(
γ
)
=
∑
t
∈
𝒯
max
{
-
m
γ
t
-
1
,
0
}
γ
t
≥
f
(
x
t
,
ω
ℓ
)
-
y
t
,
∀
t
∈
𝒯
γ
t
≥
y
t
-
f
(
x
t
,
ω
u
)
,
∀
t
∈
𝒯
∑
t
∈
𝒯
[
f
(
x
t
,
ω
u
)
-
f
(
x
t
,
ω
ℓ
)
]
≤
v
0
≤
f
(
x
t
,
ω
ℓ
)
≤
f
(
x
t
,
ω
u
)
≤
1
,
∀
t
∈
𝒯
where L vex + (γ) and L vex − (γ) are both convex functions, which constitute a minuend and a subtrahend of the difference of convex functions L DC (γ t ; m), respectively; δ (k) is a subdifferential of the convex function L vex − (γ) at γ (k) , satisfying
δ
(
k
)
∈
{
g
∈
ℝ
❘
"\[LeftBracketingBar]"
𝒯
❘
"\[RightBracketingBar]"
❘
L
vex
-
(
γ
)
≥
L
vex
-
(
γ
(
k
)
)
+
g
T
(
γ
-
γ
(
k
)
)
,
∀
γ
}
=
{
[
g
1
g
2
⋯
g
❘
"\[LeftBracketingBar]"
𝒯
❘
"\[RightBracketingBar]"
]
T
❘
∀
t
∈
𝒯
,
{
g
t
=
-
m
if
γ
t
≤
0
g
t
∈
[
-
m
,
0
]
if
γ
t
=
0
g
t
=
0
if
γ
t
>
0
}
where g ∈ | | is a real column vector with a dimensionality equal to the number of samples | | in the training set;
step (4): the iteration counter self-increasing k←k+1; calculating a convergence error e←θ (k) −θ (k-1) ; and
step (5), checking whether a Euclidean norm ∥e∥ 2 of the convergence error meets the convergence accuracy ∈ 2 , and if not, returning to the step (3), otherwise outputting a converged solution ω ←θ (k) .