Intelligent Extraction Method for Reservoir Operation Rules

Abstract

The present invention belongs to the field of flood and drought disaster prevention, water resources management and allocation, and relates to an intelligent extraction method for reservoir operation rules, comprising the following steps: 1) Classifying the optimal storage and release strategies obtained from deterministic optimal operation using a two-layer clustering method; 2) Extracting the corresponding operation rules for different operation modes; 3) Judging the most suitable operation mode to be adopted in the upcoming period using a two-step method; 4) Making a decision on the downstream discharge in the upcoming period according to the most suitable operation mode determined in step 3). The present invention provides an intelligent extraction method for reservoir operation rules that can improve the discrimination accuracy and provide technical support for the realization of the comprehensive benefits of the reservoir.

Description

Technical Field

[0001] The present invention belongs to the fields of flood and drought disaster prevention and water resources management and allocation, and relates to a method for intelligently extracting reservoir operation rules, in particular to a method for intelligently extracting reservoir operation rules based on double-layer clustering and two-step discrimination. Background Art

[0002] Reservoir operation rules are indispensable technical guidelines for guiding reservoir operation. Usually, the operation rules are extracted based on experience or by statistically analyzing the results of deterministic optimal reservoir operation (commonly known as the implicit stochastic optimization method). Deterministic optimal reservoir operation is to calculate the optimal storage and release strategies of the reservoir by using deterministic optimization methods such as dynamic programming based on historical inflow runoff sequences. Since the future inflow runoff series of the reservoir is random and non-repetitive, the results of deterministic optimal reservoir operation are often difficult to directly guide the actual reservoir operation. Therefore, it is necessary to analyze the relationship and characteristics between the historical optimal storage and release strategies of the reservoir and the historical water inflow and water demand of the reservoir, and extract a set of operation rules applicable to guiding the future reservoir operation. By mining the results of deterministic optimal reservoir operation, practical operation rules of the reservoir with guiding significance can be extracted, which can effectively enhance the practicability of theoretical optimal reservoir operation results, provide more accurate decision-making support for the actual reservoir operation, and thus maximize the comprehensive benefits of multi-functional reservoirs.

[0003] However, the traditional implicit stochastic optimization method for extracting reservoir operation rules has difficulty in accurately expressing the non-linear relationship between decision variables such as reservoir discharge and reservoir water level and the inflow. Artificial intelligence models such as artificial neural networks, random forests, and deep learning are significantly superior to the traditional implicit stochastic optimization method in describing this non-linear relationship and have achieved remarkable results in the practice of extracting reservoir operation rules. However, these artificial intelligence models have two problems that need to be improved in terms of accuracy and practicability: 1) In the extraction of operation rules, problems that violate operation common sense such as new violations of the upper and lower boundary constraints of the water level and the upper and lower limits of the discharge are likely to occur; 2) There is a problem that the generalization ability needs to be improved because the same artificial intelligence model is used for the operation rules in different operation periods such as the pre-flood drawdown period and the post-flood storage period. Summary of the Invention

[0004] In order to solve the above two technical problems in the background art, the present invention provides a new method for extracting operation rules. This method not only improves the simulation accuracy of the operation rules but also enhances the rationality of the operation rules, reduces the probability of violating reservoir operation constraint conditions, and can provide technical support for the realization of the comprehensive benefits of the reservoir.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] An intelligent extraction method for reservoir operation rules, characterized in that: the intelligent extraction method for reservoir operation rules includes the following steps:

[0007] 1) Classify the optimal storage and release strategies obtained from deterministic optimal operation using a two-layer clustering method;

[0008] 2) Extract the corresponding operation rules for different operation modes;

[0009] 3) Use a two-step method to determine the most suitable operation mode for the upcoming period;

[0010] 4) Make a decision on the downstream discharge flow for the upcoming period.

[0011] Preferably, the specific implementation method of step 1) adopted in the present invention is:

[0012] 1.1) Obtain the optimal storage and release strategies obtained from deterministic optimal operation;

[0013] 1.2) Adopt the agglomerative hierarchical clustering method, and consider the similarity of reservoir inflow and outflow to conduct a preliminary clustering of the entire optimal storage and release strategies to determine the major operation mode categories;

[0014] 1.3) Adopt the subtractive clustering method. Considering the subtle differences in the operation period, inflow, and outflow, etc., divide the major operation mode categories determined in step 1.2) into smaller subcategories, and each subcategory represents an operation mode.

[0015] Preferably, the specific implementation method of step 1.1) adopted in the present invention is to optimize based on the historical runoff sequence using mathematical programming models such as linear programming, dynamic programming, and bilevel programming. If the bilevel programming model is used, the upper layer can aim at minimizing the occupancy rate of the flood control storage capacity, and the lower layer can aim at maximizing the annual power generation. The particle swarm optimization algorithm can be used to solve the bilevel programming model to obtain the optimal storage and release strategies.

[0016] Preferably, the expression of the major operation mode category in step 1.2) adopted in the present invention is:

[0017]

[0018] Where:

[0019] Q in (t) is the reservoir inflow at time t;

[0020] Q out (t) is the reservoir outflow at time t in the optimal storage and release strategies;

[0021] AHC(k) is the k-th major operation mode category obtained by the agglomerative hierarchical clustering method;

[0022] AHC is the Agglomerative Hierarchical Clustering method;

[0023] T is the total number of time periods of the historical runoff series.

[0024] Preferably, the expression of the subclass in step 1.3) adopted by the present invention is:

[0025]

[0026] Where:

[0027] Q in (t) is the reservoir inflow at time period t;

[0028] Q out (t) is the reservoir outflow at time period t in the optimal storage and release strategy;

[0029] Z t is the reservoir water level at time period t in the optimal storage and release strategy;

[0030] D t is the specific date of time period t;

[0031] C(i, m) is the m-th scheduling mode after the subtractive clustering method subdivides AHC(i);

[0032] SCM is the subtractive clustering method.

[0033] Preferably, the scheduling modes adopted in step 1) of the present invention include AHC(1) with almost equal outflow and inflow, AHC(2) with a fixed discharge flow rate, and AHC(3) with a non-linear functional relationship between the outflow and the inflow.

[0034] Preferably, when the major scheduling mode adopted in the present invention is AHC(1) with almost equal outflow and inflow, the subclass corresponding to AHC(1) with almost equal outflow and inflow is C(1, 1), and the following expression is adopted in step 2) to extract the corresponding scheduling rules of different scheduling modes:

[0035] Q d (t) = Q in (t)

[0036] Where:

[0037] Q d (t) is the simulated reservoir outflow at time period t under the extracted scheduling mode;

[0038] Q in (t) is the reservoir inflow at time period t.

[0039] Preferably, when the major scheduling mode adopted in the present invention is AHC(2) with a fixed discharge flow rate, the AHC(2) with a fixed discharge flow rate adopts different fixed discharge flow rates according to the subcategory C(2, m), and the following expression is used in step 2) to extract the corresponding scheduling rules for different scheduling modes:

[0040] Q d (t) ∈ {Q min , Q max , Q3, …, Q n}

[0041] Where:

[0042] Q d (t) is the simulated outflow at time t under the extracted scheduling mode;

[0043] Q min is the minimum discharge flow rate of the reservoir;

[0044] Q max is the maximum discharge flow rate of the reservoir;

[0045] Q n is the fixed discharge flow rate of the reservoir to meet other functional requirements.

[0046] Preferably, when the major scheduling mode adopted in the present invention is AHC(3) where the outflow is a functional relationship of the inflow, the AHC(3) where the outflow is a functional relationship of the inflow uses the extreme learning machine for simulation for each subclass C(3, m), and the following expression is used in step 2) to extract the corresponding scheduling rules for different scheduling modes:

[0047] Q d (t) = ELM[Q in (t), D(t)]

[0048] Where:

[0049] is Q d (t) is the simulated outflow at time t under the extracted scheduling mode;

[0050] Q in (t) is the inflow of the reservoir at time t;

[0051] is the specific date of D t under the extracted scheduling mode at time t;

[0052] ELM is the extreme learning machine.

[0053] Preferably, the specific implementation manner of step 3) adopted in the present invention is:

[0054] 3.1) Use the decision tree and the hidden Markov model respectively to predict the scheduling mode to be adopted in the upcoming period;

[0055] 3.2) Use the adaptive boosting algorithm to make a final judgment on the prediction results obtained in step 3.1) to obtain the most suitable scheduling mode to be adopted in the upcoming period.

[0056] Preferably, the specific implementation method of step 4) adopted in the present invention is to calculate the discharge flow according to the information such as the reservoir inflow, reservoir water level and scheduling period in the upcoming period and the scheduling rules corresponding to the most suitable scheduling mode determined in step 3).

[0057] The advantages of the present invention are:

[0058] The present invention provides a method for intelligently extracting reservoir scheduling rules, especially a method for intelligently extracting reservoir scheduling rules based on double-layer clustering and two-step discrimination. The double-layer clustering first uses the agglomerative hierarchical clustering method and the subtractive clustering method to achieve a fine classification of the scheduling modes contained in the deterministic optimal scheduling results. The two-step discrimination method first uses the decision tree-hidden Markov model and the adaptive boosting algorithm to achieve an accurate discrimination of the scheduling mode in the upcoming period. The present invention fully considers the characteristics that the relationship between reservoir outflow and inflow and the scheduling mode are quite different due to different scheduling objectives in different scheduling periods such as the pre-flood recession period, flood season, post-flood water storage period, and dry season of the reservoir. The double-layer clustering method is used to finely classify the scheduling modes and the relationship between reservoir outflow and inflow considering different scheduling periods in the deterministic optimal scheduling results, and different models are used to extract the scheduling rules under different scheduling modes respectively. In the upcoming period during the actual scheduling process, the two-step discrimination method is used to determine the scheduling mode that the reservoir should adopt and its corresponding scheduling rules, so as to finally determine the decision of the reservoir discharge flow in the upcoming period. Compared with the single-layer clustering methods such as k-means and fuzzy clustering, the double-layer clustering method can perform a more fine classification of the scheduling modes and the relationship between reservoir outflow and inflow in different scheduling periods, and use different models to extract the scheduling rules under different scheduling modes respectively, laying a good model structure foundation for effectively improving the simulation accuracy of the scheduling rules, ensuring the scientificity and rationality of mode classification to a greater extent, fully excavating the characteristics and laws of the clustering data, being able to extract more practical scheduling rules, and providing technical support for the exertion of the comprehensive benefits of the reservoir. The two-step discrimination method for the scheduling mode in the upcoming period provided by the present invention greatly reduces the error rate of mode discrimination. Description of the Drawings

[0059] Figure 1 is the flow chart of the method for extracting reservoir scheduling rules provided by the present invention;

[0060] Figure 2 is the process schematic diagram of the double-layer clustering method adopted by the present invention for classifying the scheduling modes of the optimal water storage and release strategy;

[0061] Figure 3 It is a schematic diagram of the annual distribution of the scheduling mode categories extracted by the present invention and the optimal - simulated discharge process; Specific implementation manner

[0062] The following will, through examples and in conjunction with the accompanying drawings, elaborate in detail on the specific implementation manner of the method of the present invention.

[0063] Refer to Figure 1 , an intelligent extraction method for reservoir scheduling rules proposed by the present invention, the specific steps are as follows:

[0064] Step 1, scheduling mode classification. That is, a two - layer clustering method is used to classify the optimal storage and release strategies obtained from deterministic optimal scheduling, considering the relationship with influencing factors such as the inflow. As Figure 1 shown, the specific solution steps included in this step are as follows:

[0065] Step 1.1, establish a deterministic optimization model and solve the optimal storage and release strategy. In this embodiment, a two - layer programming (BLP) model is used to optimize the historical inflow runoff sequence. The upper layer aims to minimize the occupancy rate of the flood control storage capacity, and the lower layer aims to maximize the annual power generation. The particle swarm optimization (PSO) algorithm is used to solve this two - layer programming model.

[0066] Step 1.2, considering the similarity between the reservoir inflow and outflow, the Agglomerative Hierarchical Clustering method (AHC) is used to preliminarily classify the scheduling modes, so that they are aggregated into k major scheduling mode categories AHC(k):

[0067]

[0068] Among them, Q in (t) is the reservoir inflow at time t, Q out (t) is the reservoir outflow at time t in the optimal storage and release strategy, and AHC(k) is the k - th major scheduling mode category obtained by the Agglomerative Hierarchical Clustering method.

[0069] Usually, the reservoir contains at least 3 major scheduling mode categories, including AHC(1) where the outflow is almost equal to the inflow, AHC(2) with a fixed discharge flow, and AHC(3) where the outflow is a functional relationship of the inflow.

[0070] Step 1.3, considering the subtle differences within the major scheduling mode category AHC(k), the Subtractive Clustering Method (SCM) is used to finely classify the scheduling modes, so that each major category is aggregated into 1 - m scheduling modes:

[0071]

[0072] Among them, Z t is the reservoir water level at time t in the optimal water storage and release strategy, D t is the specific date of time period t, and C(i, m) is the m-th scheduling mode after the sub-division of AHC(i) by the subtraction clustering method.

[0073] Step 2: Extraction of scheduling rules. That is, select appropriate model algorithms according to the characteristics of each scheduling mode to extract the corresponding scheduling rules of different scheduling modes. For the three major categories of scheduling modes, the appropriate algorithms for the reservoir are as follows:

[0074] For the major scheduling mode category AHC(1) where the outflow is almost equal to the inflow, for its corresponding subclass C(1, 1), use the identity function to determine the reservoir outflow, and the specific expression is:

[0075] Q d (t) = Q in (t)

[0076] Among them: Q d (t) is the simulated reservoir outflow at time t under the extracted scheduling mode;

[0077] For the major scheduling mode category AHC(2) that adopts a fixed discharge during a specific scheduling period, its reservoir outflow adopts different fixed discharges according to the subclass C(2, m), and the specific expression is:

[0078] Q d (t) ∈ {Q min , Q max , Q3,..., Q m}

[0079] Among them: Q min is the minimum reservoir discharge, Q max is the maximum reservoir discharge, and Q m is the fixed discharge for the reservoir to meet other functional requirements;

[0080] For the major scheduling mode category AHC(3) where the outflow is a functional relationship with the inflow, for each subclass C(3, m), use the Extreme Learning Machine (ELM for short) for simulation respectively, and the specific expression is:

[0081] Q d (t) = ELM[Q in (t), D(t)]

[0082] Step 3: Discrimination of scheduling modes. That is, use a two-step method to discriminate the most suitable scheduling mode for the current time period based on information such as the inflow, water level, and scheduling period of the current time period. For exampleFigure 1 As shown in the figure, the specific solution steps included in this step are as follows:

[0083] Step 3.1: Input information such as the incoming flow, water level, and scheduling period in the facing period, as well as the classification result of the scheduling mode, into the decision tree (DT) model for a preliminary judgment of the scheduling mode;

[0084] Step 3.2: Input information such as the incoming flow, water level, and scheduling period in the facing period, as well as the classification result of the scheduling mode, into the hidden Markov model (HMM) for a preliminary judgment of the scheduling mode;

[0085] Step 3.3: Input the preliminary judgment results of Step 3.1 and Step 3.2 into the adaptive boosting algorithm (AdaBoost) for a final judgment of the scheduling mode to be adopted in the facing period, and obtain the most suitable scheduling mode for the facing period.

[0086] Step 4: Make a decision on the reservoir discharge flow. That is, based on information such as the reservoir inflow, reservoir water level, and scheduling period in the facing period, and the scheduling rules corresponding to the most suitable scheduling mode determined in Step 3.3, calculate the discharge flow.

[0087] Taking the extraction of the Three Gorges Reservoir scheduling rules as an example, the effectiveness and rationality of the inventive method are illustrated. The Three Gorges Reservoir has functions such as flood control, power generation, shipping, water resource utilization, and ecological environment protection. In actual scheduling operation, to consider the above functions as a whole, four relatively distinct scheduling periods, namely the pre-flood drawdown period, the flood season, the post-flood storage period, and the dry season, have been formed. Each scheduling period has a relatively fixed scheduling mode according to the upstream and downstream incoming water and downstream water demand conditions. The embodiment used the daily inflow runoff sequences of 30 years (20 years for training and 10 years for verification) to classify the scheduling modes of the Three Gorges Reservoir, extracted the scheduling rules under each type of scheduling mode, and calculated the corresponding discharge flow. Figure 2 Shows the clustering process of the double-layer clustering method. First, the aggregation hierarchical clustering method AHC is used to divide the policy set data into 3 major categories. On this basis, the subtractive clustering method SCM is used to further divide the policy set data into 7 types of scheduling modes. Figure 3 Intuitively shows the time distribution of the 7 types of scheduling modes in 2000 and the optimal water storage and release process under this scheduling mode. In the figure, the C(2, 1) mode is in the dry season, and the reservoir discharges while maintaining the ecological base flow; for the C(1, 1) mode and the C(1, 2) mode, the reservoir discharges according to the incoming flow to keep the reservoir water level at the normal storage level or the flood limit level; the C(2, 2) mode is in the flood season, and the reservoir discharges according to the downstream safe discharge; the C(3, 1) mode belongs to the pre-flood drawdown period, and the reservoir discharge flow exceeds the incoming flow, and the reservoir water level gradually drops; the C(3, 2) mode and the C(3, 3) mode belong to the post-flood storage period, and the reservoir discharge flow is less than the incoming flow, and the reservoir water level gradually rises. Figure 3Also shown are the optimal discharge calculated by the deterministic optimization model in 2000 and the simulated discharge under the extraction scheduling rules. As can be seen from the figure, after the scheduling mode is refined, the scheduling rules of some scheduling modes can be more accurately extracted, the data law in the deterministic optimization scheduling results is fully explored, so that the generalization ability of the model is enhanced, and the extracted scheduling rules are more in line with the actual situation. The discrimination accuracy of the two-step method in the verification period reaches 93.2%, the certainty coefficient of the discharge is 0.97, and the qualified rate of the discharge (the proportion with an absolute error less than 10%) is 97.4%.

Claims

1. An intelligent extraction method for reservoir operation rules, characterized in that: The intelligent extraction method of the reservoir operation rule includes the following steps: 1) Classify the optimal storage and release strategies obtained by deterministic optimal operation using the double-layer clustering method. The specific implementation method is as follows: 1.1) Obtain the optimal storage and release strategies obtained by deterministic optimal operation. The specific implementation method is as follows: Optimize based on the historical runoff sequence using a bilevel programming model. The upper layer aims to minimize the occupancy rate of the flood control storage capacity, and the lower layer aims to maximize the annual power generation. The optimal storage and release strategies are obtained by solving the bilevel programming model using the particle swarm algorithm; 1.2) Adopt the agglomerative hierarchical clustering method to conduct a preliminary clustering of the scheduling patterns for the entire optimal storage and release strategies considering the similarity of reservoir inflow and outflow, and determine the major scheduling pattern categories; 1.3) Adopt the subtractive clustering method to divide the major scheduling pattern categories determined in step 1.2) into smaller subcategories, and each subcategory represents a scheduling pattern; 2) Extract the corresponding operation rules for different scheduling patterns; 3) Use a two-step method to determine the most suitable scheduling pattern for the upcoming period. The specific implementation method is as follows: 3.1) Input the information of the upcoming period and the scheduling pattern classification results into the decision tree model for a preliminary judgment of the scheduling pattern; 3.2) Input the information of the upcoming period and the scheduling pattern classification results into the hidden Markov model for a preliminary judgment of the scheduling pattern; 3.3) Input the preliminary judgment results obtained in step 3.1) and step 3.2) into the adaptive boosting algorithm for a final judgment of the scheduling pattern to be adopted in the upcoming period, and obtain the most suitable scheduling pattern for the upcoming period; In step 3.1) and step 3.2), the information includes inflow, water level, and operation period; 4) Make a decision on the outflow discharge in the upcoming period. The specific implementation method is as follows: Calculate the outflow discharge according to the information of the upcoming period and the operation rules corresponding to the most suitable scheduling pattern determined in step 3); the information is the reservoir inflow, reservoir water level, and operation period.

2. The intelligent extraction method of reservoir operation rules according to claim 1, characterized in that: The expression of the major scheduling pattern category in step 1.2) is: Where: Q in (t) is the reservoir inflow in time period t; Q out (t) is the reservoir outflow at time t in the optimal storage and release strategy; AHC(k) is the major scheduling pattern category of the kth category obtained by the agglomerative hierarchical clustering method; AHC is the agglomerative hierarchical clustering method; T is the total number of periods of the historical runoff sequence.

3. The intelligent extraction method of reservoir operation rules according to claim 1, characterized in that: The expression of the subcategory in step 1.3) is: Where: Q in (t) is the reservoir inflow during time period t; Q out (t) is the reservoir outflow during time period t in the optimal water storage and release strategy; Z t is the reservoir water level at time t in the optimal water storage and release strategy; D t is the specific date of time period t; C(i, m) is the mth scheduling pattern category after the subtractive clustering method subdivides AHC(i); SCM is the subtractive clustering method.

4. The intelligent extraction method of reservoir operation rules according to claim 1, wherein: The scheduling patterns in step 1) include AHC(1) with equal outflow and inflow, AHC(2) with a fixed outflow discharge, and AHC(3) with a non-linear functional relationship between outflow and inflow.

5. The intelligent extraction method of reservoir operation rules according to claim 4, characterized in that: When the major scheduling pattern category is AHC(1) with equal outflow and inflow, the corresponding subcategory of AHC(1) with equal outflow and inflow is C(1, 1). In step 2), the following expression is used to extract the corresponding operation rules for different scheduling patterns: Q d Q(t) = in Q(t) Where: Q d (t) is the simulated outbound flow rate at time period t in the extraction scheduling mode; Q in (t) is the reservoir inflow in the t-th period.

6. The intelligent extraction method of reservoir operation rules according to claim 4, characterized in that: When the major scheduling pattern category is AHC(2) with a fixed outflow discharge, AHC(2) with a fixed outflow discharge adopts different fixed outflow discharges according to the subcategory C(2, m). In step 2), the following expression is used to extract the corresponding operation rules for different scheduling patterns: Q d (t) ∈ {Q min , Q max , Q3, …, Q n} Where: Q d (t) is the simulated outbound flow rate at time period t in the extraction scheduling mode; Q min is the minimum discharge flow of the reservoir; Q max is the maximum discharge of the reservoir; Q n is the fixed discharge of the reservoir to meet other functional requirements.

7. The intelligent extraction method of reservoir operation rules according to claim 4, characterized in that: When the major scheduling mode category is AHC(3) where the outflow is a functional relationship of the inflow, the AHC(3) with the outflow being a functional relationship of the inflow is simulated using the extreme learning machine for each subclass C(3, m), and the following expression is used in step 2) to extract the corresponding scheduling rules for different scheduling modes: Q d (t) = ELM[Q in (t), D(t)] Where: Q d (t) is the simulated outbound flow rate at time period t in the extraction scheduling mode; Q in (t) is the reservoir inflow during time period t; D t is the specific date of the t period; is the time period, is to extract the ELM (extreme learning machine) under the scheduling mode.

Images