CN105976101A - Prediction-decision making coupled reservoir operation method based on SVM (Support Vector Machine) and DPY (Dynamic Programming modified by Yang Guang) - Google Patents

Abstract

The invention discloses a prediction-decision making coupled reservoir operation method based on a SVM (Support Vector Machine) and DPY (Dynamic Programming modified by Yang Guang). An SVM runoff prediction method is adopted, a reservoir runoff process is predicted, a DPY algorithm is adopted for reservoir optimal operation, and corresponding flexible prediction-decision space is obtained. An NSGA-II algorithm is adopted, benefit and stability serve as targets, parameters n and M in the reservoir operation flexible prediction-decision model are optimized, and the benefit and the stability are checked. According to the method of the invention, the prediction information can be used effectively, the reservoir operation benefit and the decision making stability are ensured, and an important and strong-maneuverability reference basis is provided for scientific decision making for a further reservoir in a condition with runoff prediction uncertainty.

Description

A kind of forecast-decision-making based on SVM with DPY couples reservoir operation method

Technical field

The present invention relates to reservoir operation technical field, a kind of based on SVM with DPY pre- Report-decision-making coupling reservoir operation method.

Background technology

Reservoir as the engineering of a regulated flow, make water resource be more suitable for human social development and Safeguard that ecological environment aspect plays an important role.Meanwhile, hydrologic forecast and scheduling model is uncertain Property cause reservoir operation decision-making exist uncertainty.Along with human progress and social development, how in footpath There is science decision under probabilistic situation in stream forecast, plays the comprehensive benefit of reservoir to following reservoir Operational management important in inhibiting.

In recent years, along with the Runoff Forecast deep development in terms of reservoir operation field, Chinese scholars Propose Runoff Forecast method based on support vector machine (Support Vector Machine, SVM). Such as: week show equality (Zhou Xiuping, Wang Wensheng, Huang Weijun. Support vector regression model is at runoff Application [J] in prediction. HYDROELECTRIC ENERGY science, 2006,24 (4): 4-7) utilize moon runoff for many years Sequence establishes Runoff Predicting Model based on SVM, translates into the forecasting problem of annual flow by 3 The function regression problem of the SVM of individual input and 1 output.Zhang Nan etc. (Zhang Nan, the summer improves oneself, Jiang Hong. Support vector machine Runoff Forecast [J] based on multiple-factor quantizating index. Journal of Hydraulic Engineering, 2010,41 (11): 1318-1323.) utilize least square SVM method, construct runoff based on multiple-factor quantizating index Forecast model, it is achieved that Runoff Forecast precision and the unification of practicality.But, they do not have footpath Stream Forecasting Methodology is incorporated in existing reservoir operation decision-making technic.SVM exists as a kind of statistical method Can there is uncertainty during prediction runoff, reservoir operation combined with it also should be based on uncertainty Decision-making.In order to set up based on (Yang Guang, Guo such as probabilistic reservoir operation FLEXIBLE DECISION, Yang Guang Raw white silk, Li Liping. consider step reservoir FLEXIBLE DECISION research [J] of ecological flow. the Central China University of Science and Technology Journal: natural science edition, 2015,43 (9): 114-116.) for considering the step reservoir of ecological flow Scheduling proposes a kind of improved dynamic programming (the Dynamic based on " different rail is with effect " Programming modified by Yang Guang, DPY), actual application shows: the method pair Uncertain impact in decision-making has good adaptability, in the case of decision space changes Remain to generation and stablize rational result of calculation.

Although Runoff Forecast based on SVM has preferable precision, but lacks the congenital of mechanistic description Deficiency also causes model to create uncertainty；Existing reservoir operation technology the most individually considers that runoff is pre- Survey or the uncertainty of scheduling decision, two kinds of uncertainties are considered simultaneously, provide effective for policymaker The dispatching technique research of decision region is the fewest.

Summary of the invention

It is an object of the invention to the deficiency overcoming prior art to exist, it is provided that one considers that runoff is pre-simultaneously Survey uncertain with scheduling decision, it is ensured that forecast based on SVM Yu the DPY-decision-making coupling of decision-making stability Heshui storehouse dispatching method.

A kind of forecast-decision-making based on SVM with DPY of the present invention couples reservoir operation method, including as follows Step:

Step 1, gathers temperature and the data information of rainfall of each website observation of reservoir upstream；

Step 2, uses SVM Runoff Forecast method, according to each moon of each website in reservoir upstream observation Heavy rainfall, minimum rainfall and average rainfall, each moon the highest temperature, the lowest temperature and average air Temperature, sets up monthly discharge prediction model, and uses Optimization Model of Genetic Algorithm parameter, it was predicted that two Phase flow Process；

Step 3, according to predicting the two Phase flow process obtained in step 2, uses DPY algorithm to water Storehouse is optimized scheduling, obtains corresponding FLEXIBLE DECISION space；

Step 4, by actual measurement two Phase flow as input, presets in FLEXIBLE DECISION space at random The scheduling of number of times, and add up the meansigma methods of all number of times benefits, as benefit P of this decision space；

Step 5, will actual measurement two Phase flow as input, use dynamic programming to obtain reservoir operation Excellent track, and add up optimal trajectory and fall in period number Ls in FLEXIBLE DECISION space；

Step 6, falls empty in FLEXIBLE DECISION so that benefit P in step 4 is maximum with optimal trajectory in step 5 Between period number Ls be up to target, use NSGA-II algorithm to reservoir operation FLEXIBLE DECISION model In parameter n and M be optimized, and the parameter back substitution of optimization is entered in DPY algorithm, is calculated Consider Runoff Forecast and scheduling decision probabilistic FLEXIBLE DECISION space collection simultaneously.

Described step 2 includes following sub-step:

(2.1) l group training sample data (x is set_i,y_i) (i=1,2 ..., l), x_i∈Rⁿ,y_i∈R；x_iFor sample Input, including n observation project, each moon including but not limited to the observation of each website in reservoir upstream is maximum Rainfall, minimum rainfall and average rainfall, each moon the highest temperature, the lowest temperature and temperature on average； y_iExport for sample, represent the two Phase flow of prediction；

(2.2) weigh in step (2.1) and return the difference degree between f (x) and the corresponding y produced, Guarantee that when returning f (x) and the y produced and differing less than ε, error is ignored, otherwise by they differences Absolute value is as error；

(2.3) projected relationship is become the optimization problem that forecast error is minimum；

(2.4) parameter in genetic algorithm optimization SVM homing method is used, and pre-according to the parameter determined Survey two Phase flow process.

Step 3 includes following sub-step:

(3.1) to predict that the two Phase flow process obtained, for input, uses DPY algorithm to carry out preliminary election number of times Optimizing scheduling of reservoir；

(3.2) the reservoir level change procedure obtained after preset times traffic control is added up, select The maximum of operating water level and minima in each period, respectively as the change of this period reservoir level Upper and lower bound, thus constitute FLEXIBLE DECISION space.

Compared with prior art, the beneficial effects of the present invention is: Appropriate application rainfall and temperature pre- Notify breath, it is provided that consider that Runoff Forecast and scheduling decision are uncertain simultaneously, improve reservoir benefit and Robustness preferable reservoir operation new method；While improving reservoir on-road efficiency, run for reservoir Provide reliable and stable decision space, greatly facilitate the operational management of reservoir.

Accompanying drawing explanation

Fig. 1 is SVM parameter determination and prediction flow chart；

Fig. 2 is DPY algorithm flow chart；

Fig. 3 a is reservoir operation track schematic diagram；

Fig. 3 b is FLEXIBLE DECISION space schematic diagram；

Fig. 4 is FLEXIBLE DECISION space collection Optimizing Flow figure.

Detailed description of the invention

SVM Runoff Predicting Model is tied mutually by the present invention with reservoir operation FLEXIBLE DECISION model based on DPY Close, establish the reservoir dispatching system that forecast-decision-making is coupled, and use non-dominant genetic algorithm In (non-dominated sorting genetic algorithm II, NSGA-II) Optimized model Relevant parameter.

Below by embodiment, and combine accompanying drawing, the technical scheme of invention be further elaborated with:

Step 1, gathers the data information such as temperature and rainfall of reservoir upstream each website observation；

Step 2, uses SVM Runoff Forecast method, according to each moon of each website in reservoir upstream observation Heavy rainfall, minimum rainfall and average rainfall, each moon the highest temperature, the lowest temperature and average air The information such as temperature, set up monthly discharge prediction model, and use Optimization Model of Genetic Algorithm parameter, finally To the two Phase flow process of prediction, SVM parameter determination and pre-flow gauge are shown in Fig. 1.

Step 2 farther includes following sub-step:

(1) l group training sample data (x is set_i,y_i) (i=1,2 ..., l), x_i∈Rⁿ,y_i∈R。x_iFor sample Input, including n observation project: each moon maximum rainfall of each website in reservoir upstream observation, Little rainfall and average rainfall, each moon the highest temperature, the lowest temperature and temperature on average etc.；y_iFor sample This output, represents the two Phase flow of prediction in the present invention.For set up reservoir upstream observation rainfall and Contacting between temperature record and predicted two Phase flow, utilizes nonlinear mappingData are mapped For high-dimensional feature space, and in higher dimensional space, sample input and sample output are returned, its letter Number is:

In formula: weight vector ω^T∈Rⁿ,For nonlinear function, b ∈ R is bias.It is constructed such that Function f (x) for the x outside sample set, can accurately estimate corresponding y.

(2) in order to weigh, (1) returns the difference degree between f (x) and the corresponding y produced, Quote insensitiveness parameter ε build loss function:

${|y-f\left(x\right)|}_{\mathrm{\ϵ}}=\left\{\begin{array}{cc}0& if|y-f\left(x\right)|\≤\mathrm{\ϵ}\\ |y-f\left(x\right)|-\mathrm{\ϵ}& otherwise\end{array}\right.---\left(2\right)$

In formula: | y-f (x) |_εRepresent when returning f (x) and the y produced and differing less than ε, their difference Can ignore, otherwise its difference should be added up.

(3) according to structural risk minimization principle, following optimization problem is defined:

In formula: γ is penalty factor, it is used for weighing the significance level of error；e_iFor insensitive loss function Relaxation factor, be used for weigh forecast error.

(4) using Lagrange function method to be optimized formula (3), detailed process is as follows:

In formula: α is Lagrange multiplier, eliminate original variable ω and e in above formula_i, obtain radially Basic function:

K (x, y)=exp (-| | x-y | |²/2σ²) (6)

σ in RBF is core width, reflects the radius that closing of the frontier comprises.According to radially Basic function can derive the estimator of Support vector regression equation:

$f\left(x\right)=\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{\mathrm{\α}}_{i}K(x,x_i)+b---\left(7\right)$

(5), after selected kernel function, the support vector in the SVM homing method of RBF is used Only need to determine parameter penalty factor γ and core width cs.Genetic algorithm is used to determine these parameters, and according to The parameter prediction two Phase flow process determined.

Step 3, according to predicting the two Phase flow process obtained in step 2, uses DPY algorithm to water Storehouse is optimized scheduling, obtains corresponding FLEXIBLE DECISION space.

Step 3 farther includes following sub-step:

(1) to predict that the two Phase flow process obtained, for input, uses DPY algorithm to carry out default time The optimizing scheduling of reservoir of number.

In dynamic programming, represent target function with the generated energy in t stage, be shown below:

E_t[V⁽ⁱ⁾(t),V^(j)(t+1)]=P_t·M_t(i, j=1 ... M) (8)

In formula: E_t[V⁽ⁱ⁾(t),V^(j)(t+1)] represent at the beginning of the t period, last current state is respectively V⁽ⁱ⁾(t)、V^(j)(t+1) time, The generated energy in power station, M is that each state is by discrete number.

Dynamic programming is using decision-making maximum for each stage benefit as Optimal Decision-making, and recurrence equation is such as Shown in following formula:

$\left\{\begin{array}{c}{E}_t^{*}\left(V^{\left(i\right)}\right(t\left)\right)=\max \{E_t\[V^{\left(i\right)}\left(t\right),V^{\left(j\right)}(t+1)\]+E_{t+1}^{*}\left(V^{\left(j\right)}\right(t+1\left)\right)\}\\ E_{T+1}^{*}\left(V^{\left(i\right)}\right(T+1\left)\right)=0\end{array}\right.,\∀t\∈T---\left(9\right)$

In formula:Represent that t stage initial equilibrium state is V⁽ⁱ⁾Time (t) (i=1～M), by finding Optimal substrategy makes power station can obtain T～t stage maximum gross generation.

And in DPY algorithm, then randomly choose a benefit preferably decision-making as Optimal Decision-making, pass Push away equation to be shown below:.

$\left\{\begin{array}{c}{E}_t^{*}\left(V^{\left(i\right)}\right(t\left)\right)=\max \{E_t\[V^{\left(i\right)}\left(t\right),V^{\left(j\right)}(t+1)\]+E_{t+1}^{*}\left(V^{\left(j\right)}\right(t+1\left)\right)\}\\ E_{T+1}^{*}\left(V^{\left(i\right)}\right(T+1\left)\right)=0\end{array}\right.,\∀t\∈T---\left(10\right)$

In formula:Represent after f (j) (j=1～M) is sorted from big to small, in the past N (n≤M) individual f (j) randomly chooses one.

DPY algorithm flow is as shown in Figure 2.

(2) the reservoir level change procedure obtained after preset times traffic control is added up, choosing Go out maximum and the minima of operating water level in each period, change respectively as this period reservoir level Upper and lower bound.The upper and lower bound of each period is coupled together and respectively constitutes whole dispatching zone Upper and lower bound, dispatching zone upper and lower bound the scope comprised is exactly obtained FLEXIBLE DECISION space, FLEXIBLE DECISION space calculation method (schematic diagram) is as shown in Figure 3.

Step 4, will survey two Phase flow as input, the FLEXIBLE DECISION space obtained in step 3 Inside randomly choose scheduling track and carry out the scheduling of preset times, calculate reservoir on-road efficiency, and statistics takes Meansigma methods, benefit P represented as this decision space.

Step 5, to survey two Phase flow as input, uses dynamic programming to obtain reservoir operation Excellent track, and add up optimal trajectory and fall period number Ls in FLEXIBLE DECISION space in step 3.

Step 6, falls empty in FLEXIBLE DECISION so that benefit P in step 4 is maximum with optimal trajectory in step 5 Between period number Ls be up to target, use NSGA-II algorithm (note: the NSGA-II that the present invention uses Algorithm is the common technology that this area processes multi-objective optimization question, has the good speed of service and Shandong Rod) parameter n in reservoir operation FLEXIBLE DECISION model and M are optimized, notify in advance utilizing On the premise of breath carries out decision-making, ensured reservoir on-road efficiency and the noninferior solution of decision-making stability simultaneously Collection, each noninferior solution correspondence one group determines the parameters optimization in FLEXIBLE DECISION space.The parameter that will optimize Back substitution enters in DPY algorithm, is calculated and considers that Runoff Forecast is probabilistic with scheduling decision soft simultaneously Property decision space collection, stablize effective decision-making for considering that the reservoir operation of Runoff Forecast provides, flexible certainly Plan space collection Optimizing Flow is as shown in Figure 4.

Claims (3)

1. forecast-decision-making based on SVM with DPY couples reservoir operation method, it is characterised in that Comprise the steps:

Step 1, gathers temperature and the data information of rainfall of each website observation of reservoir upstream；

Step 2, uses SVM Runoff Forecast method, according to each moon of each website in reservoir upstream observation Heavy rainfall, minimum rainfall and average rainfall, each moon the highest temperature, the lowest temperature and average air Temperature, sets up monthly discharge prediction model, and uses Optimization Model of Genetic Algorithm parameter, it was predicted that two Phase flow Process；

Step 3, according to predicting the two Phase flow process obtained in step 2, uses DPY algorithm to water Storehouse is optimized scheduling, obtains corresponding FLEXIBLE DECISION space；

Step 4, by actual measurement two Phase flow as input, presets in FLEXIBLE DECISION space at random The scheduling of number of times, and add up the meansigma methods of all number of times benefits, as benefit P of this decision space；

Step 5, will actual measurement two Phase flow as input, use dynamic programming to obtain reservoir operation Excellent track, and add up optimal trajectory and fall in period number Ls in FLEXIBLE DECISION space；

Step 6, falls empty in FLEXIBLE DECISION so that benefit P in step 4 is maximum with optimal trajectory in step 5 Between period number Ls be up to target, use NSGA-II algorithm to reservoir operation FLEXIBLE DECISION model In parameter n and M be optimized, and the parameter back substitution of optimization is entered in DPY algorithm, is calculated Consider Runoff Forecast and scheduling decision probabilistic FLEXIBLE DECISION space collection simultaneously.

2. the method for claim 1, it is characterised in that described step 2 includes following sub-step Rapid:

(2.1) l group training sample data (x is set_i,y_i) (i=1,2 ..., l), x_i∈Rⁿ,y_i∈R；x_iFor sample Input, including n observation project, each moon including but not limited to the observation of each website in reservoir upstream is maximum Rainfall, minimum rainfall and average rainfall, each moon the highest temperature, the lowest temperature and temperature on average； y_iExport for sample, represent the two Phase flow of prediction；

(2.2) weigh in step (2.1) and return the difference degree between f (x) and the corresponding y produced, Guarantee that when returning f (x) and the y produced and differing less than ε, error is ignored, otherwise by they differences Absolute value is as error；

(2.3) projected relationship is become the optimization problem that forecast error is minimum；

(2.4) parameter in genetic algorithm optimization SVM homing method is used, and pre-according to the parameter determined Survey two Phase flow process.

3. the method for claim 1, it is characterised in that step 3 includes following sub-step:

(3.1) to predict that the two Phase flow process obtained, for input, uses DPY algorithm to carry out preliminary election number of times Optimizing scheduling of reservoir；

(3.2) the reservoir level change procedure obtained after preset times traffic control is added up, select The maximum of operating water level and minima in each period, respectively as the change of this period reservoir level Upper and lower bound, thus constitute FLEXIBLE DECISION space.